Standard on Automated Valuation Models (AVMs)Standard on Automated Valuation Models (AVMs) Approved...

Standard on AutomatedValuation Models (AVMs)

Approved September, 2003

International Association of Assessing Officers

The assessment standards set forth herein represent a consensus in the assessing profession and have been adoptedby the Executive Board of the International Association of Assessing Officers. The objective of these standards is toprovide a systematic means by which concerned assessing officers can improve and standardize the operation of theiroffices. The standards presented here are advisory in nature and the use of, or compliance with, such standards ispurely voluntary. If any portion of these standards is found to be in conflict with the Uniform Standards ofProfessional Appraisal Practice (USPAP) or state laws, USPAP and state laws shall govern.

Published by

International Association of Assessing Officers130 East RandolphSuite 850Chicago, IL 60601-6217312/819-6100Fax: 312/819-6149http://www.iaao.org

ISBN 0-88329-180-0

Copyright © 2003 by the International Association of Assessing Officers

All rights reserved.

No part of this publication may be reproduced in any form, in an electronic retrieval system or otherwise, without the prior writtenpermission of the publisher. However, assessors wishing to use this standard for educating legislators and policymakers may photocopyit for limited distribution.

Printed in the United States of America.

AcknowledgmentsThe AVM Standard was reviewed and completed through the dedicated efforts of an Ad Hoc committeecomprising Alan S. Dornfest, AAS, Chair, Larry J. Clark, CAE, Robert J. Gloudemans, Michael W. Ireland,CAE, Patrick M. O’Connor, and William M. Wadsworth. The Committee worked closely with Nancy C.Tomberlin, who was chair of the Technical Standards Committee at that time.

Special thanks and appreciation also go to the many individuals who served as reviewers for this standard:

Richard AlmyRichard A. BorstMan ChoJohn S. CirincioneRobert C. DenneBrian G. GuerinM. Steven KaneJosephine Lim

Mark R. Linne, CAEWayne D. Llewellyn, CAEIan W. McClungJohn F. Ryan, CAERonald J. SchultzRuss ThimganJames F. Todora, CAERobert Walker

Contents1. Scope ............................................................................................................................................ 52. Introduction ................................................................................................................................. 5

2.1. Definition and Purpose of an AVM .................................................................................... 52.1.1 Definition ................................................................................................................. 52.1.2 Purpose .................................................................................................................... 52.1.3 Applicability ............................................................................................................. 52.1.4 Distinction from Traditional Valuation Applications .............................................. 5

2.2 Purpose and Use of AVMs .................................................................................................. 62.2.1 General .................................................................................................................... 62.2.2 Analysis of Impaired Properties .............................................................................. 6

2.3 Steps in AVM Development and Application ...................................................................... 62.3.1 Property Identification ............................................................................................. 62.3.2 Assumptions ............................................................................................................. 62.3.3 Data Management and Quality Analysis ................................................................. 62.3.4 Model Specification .................................................................................................. 72.3.5 Model Calibration .................................................................................................... 72.3.6 Model Testing and Quality Assurance .................................................................... 72.3.7 Model Application and Value Review ....................................................................... 72.3.8 Stratification ............................................................................................................ 72.3.9 Value Defense .......................................................................................................... 8

3. Specification of AVM Models ....................................................................................................... 83.1 Data Quality Assurance ..................................................................................................... 83.2 Model Specification Methods ............................................................................................ 9

3.2.1 Cost Approach .......................................................................................................... 93.2.2 Sales Comparison Approach .................................................................................... 9

3.2.2.1 Comparable Sales Method ............................................................................. 93.2.2.2 Direct Market Method ................................................................................... 9

3.2.3 Income Approach ....................................................................................................103.3 Stratification ...................................................................................................................103.4 Location ............................................................................................................................10

4. Calibration Techniques ............................................................................................................. 114.1 Calibration Using Multiple Regression Analysis (MRA) .............................................. 11

4.1.1 MRA Assumptions ................................................................................................ 114.1.2 Diagnostic Measures of Goodness-of-Fit ...............................................................124.1.3 MRA Software, Options and Techniques ...............................................................124.1.4 MRA Strengths ......................................................................................................124.1.5 MRA Weaknesses ..................................................................................................12

4.2 Calibrating Using Adaptive Estimation Procedure (AEP) ..............................................124.2.1 AEP Model Structure .............................................................................................134.2.2 Variable Control in AEP .........................................................................................134.2.3 Results and Goodness-of-Fit Measures .................................................................134.2.4 AEP Advantages ......................................................................................................144.2.5 AEP Disadvantages .................................................................................................14

4.3 Artificial Neural Networks ..............................................................................................144.3.1 The Artificial Neuron ............................................................................................144.3.2 Strengths of Neural Networks ..............................................................................144.3.3 Weakness of Neural Networks ..............................................................................14

4.4 Time Series Analysis ......................................................................................................154.5 Tax Assessed Value Model ..............................................................................................154.6 Calibration Summary .....................................................................................................16

5. Residential AVMs ......................................................................................................................165.1 Detached Single-Family ..................................................................................................16

5.1.1 Cost Models ............................................................................................................165.1.2 Comparable Sales Models ......................................................................................175.1.3 Direct Market Models ...........................................................................................17

5.2 Attached Residential Property (Condominiums, Townhouses, Zero-Lot-Lines) ...........175.3 Two- to Four-Family Residential Property .....................................................................185.4 Manufactured Housing ...................................................................................................185.5 Time Series Models for Residential Property ...............................................................185.6 Summary and Conclusions for Using Residential AVMs ...............................................19

4

6. Commercial and Industrial AVMs .............................................................................................196.1 Commercial and Industrial Model Specification ............................................................19

6.1.1 Property Use ..........................................................................................................196.1.2 Location ..................................................................................................................196.1.3 Physical Characteristics and Site Influences .......................................................206.1.4 Income Data ............................................................................................................20

6.2 Development of the Model(s) ...........................................................................................206.2.1 Cost Models ............................................................................................................216.2.2 Sales Comparison Models .....................................................................................216.2.3 Income Models .......................................................................................................21

6.2.3.1 Modeling Gross Income ...............................................................................216.2.3.2 Vacancy and Collection Losses ...................................................................216.2.3.3 Modeling Expenses ......................................................................................216.2.3.4 Direct Capitalization ...................................................................................216.2.3.5 Gross Income Multiplier .............................................................................226.2.3.6 Property Taxes .............................................................................................22

6.3 Quality Assurance ..........................................................................................................227. Land Models ..............................................................................................................................22

7.1 Land Valuation Model Specification ...............................................................................227.1.1 Property Use ..........................................................................................................227.1.2 Location ..................................................................................................................227.1.3 Physical Characteristics and Site Influences .......................................................23

7.2 Land Data Collection ......................................................................................................237.3 Development of the Model(s) ...........................................................................................23

7.3.1 Land Valuation Modeling by Sales Comparison ....................................................237.3.2 Land Valuation Modeling by Income ......................................................................24

8. Automated Valuation Model Testing and Quality Assurance ...................................................248.1 Data Quality Assurance ...................................................................................................248.2 Data Representativeness ..................................................................................................248.3 Model Diagnostics ............................................................................................................248.4 Sales Ratio Analysis ........................................................................................................25

8.4.1 Measures of Appraisal Level .................................................................................258.4.2 Measures of Variability .........................................................................................25

8.4.2.1 Coefficient of Dispersion .............................................................................258.4.2.2 Coefficient of Variation ...............................................................................25

8.4.3 Measures of Reliability ..........................................................................................258.4.4 Vertical Inequities .................................................................................................278.4.5 Guidelines for Evaluation of Quality .....................................................................278.4.6 Importance of Sample Size ....................................................................................27

8.5 Property Identification .....................................................................................................288.6 Outliers ............................................................................................................................298.7 Holdout Samples ...............................................................................................................298.8 Value Reconciliation ........................................................................................................298.9 Appraiser Assisted AVMs ................................................................................................308.10 Frequency of Updates .....................................................................................................30

9. AVM Reports ..............................................................................................................................309.1 Types of Reports ...............................................................................................................30

9.1.1 Documentation Report ...........................................................................................309.1.2 Restricted Use Report ............................................................................................309.1.3 CAMA or AAAVM Report ......................................................................................30

9.2 Uses of AVM .....................................................................................................................319.2.1 Real Estate Lenders ...............................................................................................319.2.2 Real Estate Professional ........................................................................................319.2.3 Government ............................................................................................................319.2.4 General Public .......................................................................................................31

Glossary ........................................................................................................................................31References ....................................................................................................................................35Additional Suggested Readings ....................................................................................................35

Standard on Automated Valuation Models (AVMs)

1. SCOPEThis standard is intended to provide guidance for bothpublic sector CAMA and private sector AVM systems.This standard provides recommendations and guidelineson the design, preparation, interpretation, and use ofautomated valuation models (AVMs) for the appraisal ofproperty. The standard presents market analysis basedappraisal applications and aspects of such models. Theprinciples addressed in this standard are consideredapplicable to all appraisals of real property, which aredesigned to estimate market value.

The standard does not address appraisal of personalproperty, such as machinery and equipment, and AVMsare not considered applicable for appraisal of highlyspecialized or unique property.

As presented in this standard, the development of an AVMconforms to USPAP Standard 6 (Appraisal Foundation2003, 46–56). The appraiser using AVM output shouldfollow USPAP standards that relate to their assignment.

2. INTRODUCTION2.1 Definition and Purpose of an AVM2.1.1 DefinitionAn automated valuation model (AVM) is a mathemati-cally based computer software program that producesan estimate of market value based on market analysis oflocation, market conditions, and real estate characteris-tics from information that was previously and separatelycollected. The distinguishing feature of an AVM is thatit is an estimate of market value produced throughmathematical modeling. Credibility of an AVM is depen-dent on the data used and the skills of the modelerproducing the AVM.

2.1.2 PurposeThe purpose of an AVM is to provide a credible, reliable,and cost-effective estimate of market value as of a givenpoint in time. Market value is the most probable price (interms of money) that a property should bring in acompetitive and open market under the conditions req-uisite to a fair sale—the buyer and seller each actingprudently and knowledgeably, and assuming the price isnot affected by undue stimulus. AVM values reviewedfor reliability, and generated in compliance with USPAPStandard 6 are considered appraisals.

AVMs are developed and used by both the public andprivate sector. Assessment officials use AVMs to pro-duce estimates of value as of a common date forpurposes of property assessment and taxation. Privatesector appraisers and their clients use AVMs to estimatethe value of a subject property at a given point in time fora wide variety of purposes.

2.1.3 ApplicabilityAVMs are applicable to any type of property for whichadequate market information and property data areavailable in the relevant market area. The relevant marketarea is the area that would be considered by potentialpurchasers. For residential properties, this is typically allor a portion of a metropolitan area, one or more townsin a geographic area, or a given rural or recreational area.The market area for larger multi-family, commercial,and industrial properties can be regional or even nationalin scope, depending on the relevant investors and marketparticipants.

The development of an AVM is an exercise in theapplication of mass appraisal principles and techniques,in which data are analyzed for a sample of properties todevelop a model that can be applied to similar propertiesof the same type in the same market area. These may beeither individual properties of interest or all propertiesthat meet the requirements of the model.

Although the same underlying principles are applicable to allAVMs, the specific formulation and calibration techniqueswill vary with the purpose of the AVM, type of property,available data, and experience and preferences of themarket analyst. Sections 3 and 4 discuss the generalprinciples of model specification and calibration. Section 5addresses residential AVMs. Section 6 focuses on com-mercial and industrial AVMs and section 7 focuses onAVMs developed for vacant or improved land.

2.1.4 Distinction from Traditional ValuationApplications

Although AVM development requires skilled analysis andattention to quality assurance, AVMs are characterized bythe use and application of statistical and mathematicaltechniques. This distinguishes them from traditional ap-praisal methods in which an appraiser physically inspectsproperties and relies more on experience and judgment toanalyze real estate data and develop an estimate of marketvalue. Provided that the analysis is sound and consistent

STANDARD ON AUTOMATED VALUATION MODELS (AVMS)—2003

6

with accepted appraisal theory, an advantage to AVMs is theobjectivity and efficiency of the resulting value estimates.Of course, sound judgment is required in model develop-ment and an appraiser should review the values producedby the model.

2.2 Purpose and Use of AVMs2.2.1 GeneralAVMs are used to provide estimates of market value fora variety of public and private sector purposes. AVMestimates reflect a given time period and should becalibrated to produce market values as of a specific date.Although past market trends can be projected over ashort time horizon, the credibility of appraisal estimatesincreasingly suffers as the projection is lengthened.

AVMs have the advantage of objectivity and consis-tency, reduced cost, and faster delivery time. It isimportant, however, that the AVM follow soundstatistical and mathematical modeling practices andbe tested for accuracy and uniformity before applica-tion. Section 8 discusses the important area of modeltesting and quality assurance and section 9 focuses onreporting of results.

2.2.2 Analysis of Impaired PropertiesProperties subject to significant defects or that areaffected by atypical circumstances impairing marketvalue, including superadequacy or functional obsoles-cence, cannot be accurately modeled with an AVM. Anappraiser may choose to apply the AVM to the property,but the defect or unique circumstance should be notedand a special adjustment made to compensate for thedefect or special circumstance.

2.3 Steps in AVM Development andApplication

The remaining portion of this section outlines the stepsto take in development of an AVM. The followingsections of this standard provide clarification and detailsconcerning these steps and their application to particularproperty types.

2.3.1 Property IdentificationThe first step in any appraisal problem is to identifythe property to be appraised. In developed econo-mies, identification is normally straightforward, asmaps, ownership records, property addresses, andlegal descriptions will identify the property andowner. The appraisal assignment will usually requireidentifying physical characteristics and propertyrights to be valued as of the appraisal date. Whenapplying an AVM to a particular property, improve-ments and renovations made before this date shouldbe included in the appraisal; those made subsequentto the appraisal date should not.

The bundle of rights to be appraised generally in-cludes the fee simple interest or full bundle of rightsinherent in ownership of property. Nevertheless, themarket analyst should make clear what rights areassumed and any limitations to full use or restrictionsto transfer of the property.

2.3.2 AssumptionsThe AVM supporting documentation should state allassumptions, special limiting conditions, extraordinaryassumptions, and hypothetical conditions. A key as-sumption in many AVM applications concerns theassumed use of the property. Most real estate databasescontain the actual use of property as of the inspectiondate. In some property tax systems, current use isstipulated as the basis for valuation. However, compa-rable market sales reflect the concept of highest and best(most probable) use. Market analysts and users ofAVMs need to be aware of these subtleties.

Another key assumption relates to whether or not the feesimple bundle of rights is being appraised. This isgenerally the case for residential properties, but manycommercial appraisals are made to estimate only theleased fee or leasehold interest when there is an existinglease (or leases) on the property.

Government appraisal agencies are responsible for collect-ing and maintaining property databases, although they oftencontract with private vendors for this purpose. Commer-cial AVM providers generally use data maintained by agovernment agency or third party service. In all cases, it isimperative that AVM market analysts test the reliability ofthe data and clearly state assumptions concerning itsaccuracy. If data important to value estimation are missingor the statistical process has shown the data to be inconsis-tent or unreliable, the AVM provider has a responsibility tonot provide a potentially misleading value estimate to theintended user.

2.3.3 Data Management and Quality AnalysisThe reliability of any appraisal depends on accurate data.Appraisal data fall into two general categories: propertydata and market data. Property data relate to location,land characteristics, and building features. Market datainclude sales, income, and cost information. Askingprices and independent appraisals can sometimes beused to supplement sparse sales data.

Computerized statistical tools used to develop AVMsafford the opportunity to screen data for missing or out-of-range occurrences and inconsistencies; examplesinclude homes with more than two fireplaces or a bi-level home with no listed lower level living area.

Geographic information systems (GIS) can also help indata reviews. GIS software is used to maintain comput-erized maps and provide geographic representations ofproperty attributes and features. It can be used to

7


highlight properties with impossible, unlikely, or incon-sistent data. For example, properties coded as beingwaterfront can be color-coded, displayed on a map, andreviewed for accuracy.

Only valid, open market sale and income data should beused in model development. (As mentioned, askingprices and independent appraisals can sometimes also beused to bolster sample sizes.)

Since the reliability of an AVM is dependent on the datafrom which it is generated, the integrity of the databaseshould be monitored on a systematic and ongoing basis.

2.3.4 Model SpecificationModel specification is the important process of determiningthe format (model structure) of the AVM. The marketanalyst must determine the type of model to be employedand specify the variables to be used in the model.

AVMs that employ property features, often character-ized as “hedonic” models, can be categorized as additive,multiplicative, or hybrid models (see Section 3 onSpecification of AVM Models). Market analysts mustalso determine the variables to be included in hedonicAVMs. These can represent property characteristics(e.g., square feet of living area and building age),location information, demographic data (e.g., incomelevels or school quality), or variables derived fromproperty characteristics (e.g., the square root of lot sizeor living area multiplied by a quality index). The objectiveis always to include property features important in valuedetermination and to capture actual market relation-ships. Skilled analysis is required to adequately specifyan effective model structure.

Some models that are referred to as AVMs have only atime component; in other words, they merely trackchanges in property values over time. Where propertycharacteristic information is unavailable or limited, thesemodels can be used to trend a previous sale or valueestimate to the target appraisal date.

2.3.5 Model CalibrationCalibration is the process of determining the coefficients inan AVM as well as which variables should be retained ordeleted due to statistical insignificance. Several statisticaltools can be used to calibrate AVM models (see Section 4on Calibration Techniques). Proper use of these toolsrequires experience and training in statistical analysis andthe software employed.

2.3.6 Model Testing and Quality AssuranceAn AVM must be tested to ensure that it meets requiredaccuracy standards before being deployed. This is accom-plished through statistical diagnostics and a ratio study inwhich value estimates (e.g., estimated sale price or esti-mated rent) are compared to actual values (e.g., sale priceor reported rent) for the same properties. GIS can be used

to display color-coded ratios on maps and help spot groupsof under- or over-valued properties. For more information,see Section 8 on Automated Valuation Model Testing andQuality Assurance. Before it is implemented, the AVM alsoshould be tested on a holdout sample, which is a set ofproperties and their selling prices that were not used in thecalibration process.

Properties with unusually large errors, termed “outliers,"should be reviewed. It is likely that the sale price (or othervalue serving as the dependent variable in the model) is notrepresentative, the data are partially incorrect, or the prop-erty exhibits atypical features that cannot be adequatelyaccounted for in the model. Except where the data can becorrected, the property should be removed from thesample, and it and similar properties with similar featuresshould not be valued by the AVM alone.

2.3.7 Model Application and Value ReviewOnce tested and validated, the AVM can be applied toestimate the value of other properties of the same typein the area or region where the model applies. Thesevalues should be reviewed for reasonableness and con-sistency with recent sales, either of the subject propertyitself or of similar properties in the same neighborhoodor surrounding area, or where sales are not available,recent asking prices.

It is also good practice to systematically review thegenerated values for reasonableness and consistencywith nearby properties in the same neighborhood. Thisaffords the opportunity to ensure that the data areaccurate, and to make individual adjustments to proper-ties with unique features or that are subject to specialinfluences, such as being located at a busy intersectionor having a premium or obstructed view.

2.3.8 StratificationStratification is the process of grouping properties formodeling and analysis. Stratification begins with prop-erty type. Properties are delineated into generic usecategories such as: single-family residential, condo-minium (if applicable), multi-family, commercial, andindustrial. The number of property types will depend onthe size and diversity of the geographic area beinganalyzed and the number of sales available within theproposed strata.

Residential properties in urban areas are generally strati-fied into “market areas.” Market areas are broad,somewhat homogeneous geoeconomic areas that appealto buyers in similar economic brackets. One AVM maybe developed for each market area, or a regional modelmay be developed and individually calibrated for eachmarket area. Location within the market area can behandled through neighborhood variables or other vari-ables related to geographic location and desirability.Alternatively, a location value response surface analysis


8

(LVRSA) may be used to measure and adjust for locationwithin the model formula (see Section 3.4 Location).

Commercial properties are usually modeled across awider geographic area than residential. For example, onemodel may be sufficient for all properties of a given type(e.g., office, retail, or warehouse) in an entire urbancounty or metropolitan area.

2.3.9 Value DefenseMarket analysts must be prepared to review and defendvalues developed through AVMs. The review processbegins with checking the accuracy of the data. If noproblem is found, the estimated value should be evaluatedfor consistency with similar properties and with any recentsales of the subject property or similar properties. The factthat a property sold for a price different from the AVMestimate does not mean that the AVM estimate is wrong.The sale date may differ significantly from the appraisaldate, and the property may have sold for a relatively low orhigh price, depending on the peculiarities of the situation andmotivations of the buyer and seller. If the estimated valueappears to be unreasonable or inconsistent with marketevidence, the AVM estimate is not reliable and should bediscarded or adjusted (the reason for the breakdown shouldbe investigated and corrected). If the estimate is supportedby market evidence, then it should be defended.

The best support for an AVM value is recent sales ofcomparable properties. Current listings can also be used,although they must be given less credence than consum-mated sales. For income properties, it may be possible tosupport a value estimate derived from one AVM (say, asales comparison model) with estimates derived fromalternative methods (e.g., an income model). The consis-tency of the value estimate with others produced by theAVM model, as well as the overall reliability of the AVMmodel as evidenced by a ratio study of the holdout sampleor other statistical measures can also be evaluated and usedto defend the value.

AVM developers should prepare documentation that willallow clients and other appraisers to understand in non-technical terms how the model was developed and applied.

3. SPECIFICATION OF AVM MODELSThe two major components of valuation are specificationand calibration. Model specification is the process ofdeveloping the proposed model structure. Model calibra-tion relates to testing the specified model structure usingdata sets to generate the model variable coefficients.

In practice the specification and calibration are performedin an iterative process which includes the following steps:

1. Specify a model

2. Test the specification with calibration

3. Make adjustments to model specification

4. Test new specification with calibration

5. Continue to repeat the process until statis-tically significant improvement is minimized

The AVM specification and calibration iterative processmakes the assumption that data are collected and verifiedin a consistent and professional manner.

3.1 Data Quality AssuranceThe model specification process begins with an evaluationof the data availability. The availability of data will influencethe specification of the model and may indicate the need forrevisions in the specification and/or limit the usefulness ofthe resulting value estimates. Publicly available data fromgovernment sources, such as government assessors, deedrecorders, registrars and census agencies, are the basis formost statistical models. Commercial sector informationservices may be used to supplement that data. Becausemore than one source will provide information toward theAVM model process, the AVM market analyst must usestatistical data analysis to confirm the assumption that thequality of the data will provide reasonable support for themodeling process.

AVM models are based on a sample of the universe ofdata. The specification process must review the sampledata used to develop the model as well as the populationto which the model will be applied. The sample shouldbe representative of the population in all key elements ofvalue including the types of properties, market condi-tions, value range, land and building sizes, and buildingages. Property types where market information is notavailable, should be excluded from both the sample andtotal population files as the model specification will notbe representative of these properties.

Indicators of value may include sale prices, rents,expenses, and capitalization rates. Limitations in theintegrity and availability of the data are important deter-minants of the model specification. Knowledge of keyproperty characteristics is crucial to model specifica-tion. Models should not be specified without anunderstanding of the data in the sample and population.

Data field verification is common in public, but not incommercial, AVM development. Commercial AVMmarket analysts rely on the accuracy of the data pro-vided to them. In cases where AVM data is not fieldverified, data quality can only be measured by its typicalrelationship to the value. When data items that apprais-ers would consider highly correlated to value do notprove to have such a relationship (correlation matrix orregression T or F values), this could be an indication ofinconsistent data collection or scarcity of data. Data thatare not consistently collected or that are mostly missingfrom the population should not be used in the model

9


specification or calibration phases, as it can be insignifi-cant and may produce misleading results.

Data may be qualitative or quantitative. Quantitative datais objective and can be counted or measured. Qualitativedata is usually descriptive, subjective, and subject tojudgmental decisions that require experience by theperson collecting the data.

3.2 Model Specification MethodsAVM models are based upon one or more of the threeapproaches to value (cost, sales comparison, and income).

3.2.1 Cost ApproachModel specification for the cost approach requires theestimation of separate land and building values.

The cost approach formula converts to a modelspecification:

MV =πGQ * [(1-BQD) * RCN + LV]

• MV is the market value estimate.

• πGQ represents the general qualitative variablessuch as location and time;

• BQD is a building qualitative variable

representing depreciation;

• RCN is the replacement/reproduction cost new;

• LV is the land value; and

(Gloudemans 1999, 124.)

If a third party provides the cost tables, it is theresponsibility of the AVM market analyst to calibrate thecost tables to the local market in order to provide a validindicator of value by the cost approach.

3.2.2 Sales Comparison ApproachThe sales comparison approach can involve either a two-step process, in which comparable sales are identified andadjusted to the subject property, or the specification andcalibration of a direct sales comparison model.

3.2.2.1 Comparable Sales MethodIn the two-step process (also referred to as the “appraisalemulation” method), one model is developed to identifycomparable sales and a second model is developed to makeadjustments for differences between the subject propertyand the identified comparables. The first model will includedata items important in determining comparability and mayinvolve the calculation of a dissimilarity measure, such asthe Minkowski or Euclidean metrics. A second model willinclude data items significant in directly estimating valuefrom the market and is used to adjust the selected compa-rable sales to the subject. Model specification for thecomparable sales method can be summarized as follows:

MVS = SPC + ADJC

• MVS represents the market value estimate;

• SPC

represents the selling price of a comparablesale property; and

• ADJC represents adjustments to the comparable

sale.


3.2.2.2 Direct Market MethodThe direct market method involves specification andcalibration of a single model to predict value directly.The model may take one of three forms: additive (alsotermed “linear”), multiplicative, or hybrid (also termed“nonlinear’). Basically, in an additive model, the con-tribution of each variable in the model is addedtogether. In a multiplicative model, the contributionsare multiplied. Hybrid models can accommodate bothadditive and multiplicative components. The choiceof model specification usually depends on the priorexperience of the market analyst and the type ofproperty being appraised. Additive models are themost prevalent of the three, based on tradition andwide availability of software programs. Nonlinear(hybrid) models are used the least due to limitedsoftware availability, but these models more accu-rately reflect the combination of additive andmultiplicative relationships in the real estate market.

Additive models have the form:

MV = B0 + B1*X1 + B2*X2 + …

• MV is the dependent variable;

• B0 is a constant;

• Xi represents the independent variables in the

model; and

• Bi are corresponding rates or “coefficients.”

In a direct sales comparison model, “MV” iseither sale price or sale price per unit. In anincome model, the dependent variable is incomeor income per unit. Additive models are relativelyeasy to calibrate and understand.

In a multiplicative model the contribution of the vari-ables is multiplied rather than added:

MV = B0 * X1B1 * X2

B2 * …

In this example each variable is raised to a correspond-ing power. However, the process can also be reversedas illustrated by the third variable in the equation below:

MV = B0 * X1B1 * X2

B2 * B3

X3 …


10

Multiplicative models consist of a base rate (B0) and

percentage adjustments. They have several advantages,including the ability to capture curvilinear relationshipsmore effectively and the ability to make adjustmentsproportionate to the value of the property being ap-praised. Multiplicative models are usually calibratedusing linear regression packages. This requires some ofthe variables to be converted to logarithmic format forcalibration, which can complicate model developmentand application.

Hybrid (nonlinear) models are a combination of additiveand multiplicative models. As such, they are theoreti-cally the best alternative of the three, but software isrelatively limited.

A general hybrid model specification that separatesvalue into building, land, and “other” components (e.g.,outbuildings) is:

MV =πGQ * [π BQ *ΣΣΣΣΣBA) +πLQ * ΣΣΣΣΣLA) + ΣΣΣΣΣOA]

• MV is the estimated market value;

• πGQ is the product of general qualitativevariables;

• πBQ is the product of building qualitativevariables;

• ΣBA is the sum of building additive variables;

• πLQ is the product of land qualitative variables;

• ΣLA is the sum of land additive variables; and

• ΣOA is the sum of other additive variables.

(IAAO 1990, 351; Gloudemans 1999, 124.)

3.2.3 Income ApproachIncome-producing real property is usually purchased forthe right to receive future income. The appraiser evaluatesthis income for quantity, quality, direction, and duration andthen converts it by means of an appropriate capitalizationrate into an expression of present worth: market value. Ifexpense data are available, the steps in this approach are:

1. Estimate gross income, expenses, and netincome from market data.

2. Select the appropriate capitalization method(model specification).

3. Estimate a capitalization rate or incomemultiplier (model calibration).

4. Compute value by capitalization.

(IAAO 2002.)

While there are many model specifications of the incomeapproach, the basic overall direct capitalization formula is:

MV = NOI/R

• MV is the price examined in the calibration andresulting estimate of market value;

• NOI is the net operating income; and

• R is the overall capitalization rate.

Another income approach methodology uses grossincome multipliers (GIMs):

MV = GI * GIM

• MV is the price examined in the calibration andresulting estimate of market value;

• GI is the gross annual income; and

• GIM is the gross income multiplier.

Gross rent multipliers are the same as gross incomemultipliers but relate to monthly gross incomes.

3.3 StratificationIn stratification, parcels are sorted into relatively homoge-neous groups based on use, physical characteristics, orlocation. Properties are first stratified by use such asagricultural, apartments, commercial, industrial, and resi-dential. Additional stratification by physical characteristicsor value ranges may be performed to minimize the differ-ences within strata and maximize differences among strata.Geographic stratification is appropriate wherever the valueof various property attributes varies significantly amongareas and is particularly effective when housing types andstyles are relatively uniform within areas (IAAO 1990, 119).Location stratification reduces the need for complex mod-els. However, excessive stratification may provide too littlevariation in the data.

When the market for a given type of property is nationalin scope, it may be possible to create national valuationmodels without stratification if location adjustments areincluded as part of the model specification and calibra-tion processes.

3.4 LocationLocation is the numerical or other identification of apoint (or object) sufficiently precise so the point can besituated. Location has a major influence upon propertyvalue. Location analysis can be used to measure therelative impact on value from the neighborhood leveldown to the individual property level. Location influ-ences within a given model area can be measured byincluding location variables in the model, or can beestablished through an analysis of the residuals (errors)from a model developed without location factors.

Two specific methods to develop location adjustmentsare the creation or use of existing neighborhoods andLVRSA. Neighborhoods are the traditional and most

11


common form of location analysis. In AVMs, neighbor-hoods may be based upon streets and natural boundaries,government assessor-designated areas, census tracts,or postal delivery codes. LVRSA techniques relaterelative prices as measured, for example, by the ratio ofeach sale price to the median price to each property’sunique location, as represented by its geographicalcoordinates. Software that provide the ability to performLVRSA use a variety of smoothing techniques to com-pute a unique location adjustment, termed the relativelocation value (RLV), for each property. At a moresophisticated level, residuals from a first model devel-oped without location variables can be plotted andanalyzed to create the RLV grid. This variable is thenincluded, along with other variables, in a multiple regres-sion or other model to capture location influences.

4. CALIBRATION TECHNIQUESModel Calibration is the development of the adjustments orcoefficients through market analysis of the variables to beused in an AVM. The definition of an AVM used in thisstandard, emphasizes the use of statistical models andprocedures in the development of the AVM. The majorityof AVMs in use rely strictly on statistical models as themethod of calibration, however USPAP Standard 6 (Ap-praisal Foundation 2003, 46–56) provides recognition ofother acceptable methods.

Multiple linear regression and nonlinear regression areclearly based in statistics, while adaptive estimationprocedure is based on a tracking method from theengineering sciences. Neural networks emulate some ofthe observed properties of biological nervous systemsand draw on the analogies of adaptive biological learning.Artificial neural networks are collections of mathemati-cal models that can emulate some of the observedproperties found in the real estate market.

4.1 Calibration Using Multiple RegressionAnalysis (MRA)

MRA is a statistically based analysis that evaluates thelinear relationship between a dependent (response) vari-able and several independent (predictor) variables, andextracts parameter estimates for independent variablesused collectively to estimate value in a mathematicalmodel. Models produced using MRA come with a richset of diagnostic statistics that provide evaluation toolsfor the market analyst to compare results between andamong specified models. These goodness-of-fit statis-tics provide information about each variable’ssignificance in predicting value, and how well thevariables in the model work together to produce credit-able results overall. Users of AVMs should be familiarwith the key measures of goodness-of-fit, and reviewthem before accepting AVM results generated by theMRA process.

4.1.1 MRA AssumptionsThe accuracy and credibility of an MRA model dependon the degree to which certain assumptions are met. Themost important assumptions are complete and accuratedata, linearity, additivity, normal distribution of errors,constant variance of the errors, uncorrelated indepen-dent variables, and sample representativeness.

Complete and accurate data is required if MRA is toachieve predictive accuracy.

Linearity assumes the marginal contribution to value by anindependent variable is constant over the entire range of thevariable. When additive models are used, this assumptionmay not be supported in the market place, requiring atransformation of the variable. Additivity continues with theconcept of marginal contribution in that any one indepen-dent variable is unaffected by the other variables in themodel. In other words, linear additive models do notpossess the ability to measure nonlinear effects or interac-tive effects of market conditions, without transforming rawvariables. In such cases, one must consider using nonlinearor hybrid models.

Normal distribution of errors follows the assumptionthat the data are normally distributed, and therefore,any error in predictions is also normally distributed.Without the assumption of the normally distributederrors, the inferences for using the standard error ofestimate and coefficient of variation (COV) as ameasure for goodness of fit are meaningless. Con-stant variance of the error term implies that theresiduals are uncorrelated with the dependent vari-able, which is the sale price. In other words, as theprice level changes, the error term remains constantor homoscedastic; when unequal variances occur atdifferent price ranges, it is heteroscedastic.

A term known as multicollinearity describes thecondition where independent variables are correlated(measure the same thing) with each other. Dependingon the method used, regression may reject one vari-able as insignificant or exaggerate coefficients forboth variables, if multicollinearity is introduced intothe model. A correlation matrix is a good tool whentesting for multicollinearity.

It is assumed the sold properties data from whichmodels are constructed are representative of the prop-erties to which they are applied. It is important that bothlow and high value properties be represented in themodel. Data should also be divided into training samplesused to develop the model, and holdout samples (controlsamples) used to test model results.

Because of its robust character, minor violation of thisassumption will not dramatically impact results. Poordata quality or samples not representative of the popu-lation will produce poor performing models.


12

The market analyst must be able to present the MRAresults in an understandable and defensible format thatappraisers and AVM clients can easily understand.

To avoid seriously violating assumption of linearity,additivity, and constant variance of the error term, themarket analyst must consider the use of transformingvariables or other calibration methods described in thestandard. A multiplicative, nonlinear, or hybrid modelstructure is best for measuring interactive effects.

4.1.2 Diagnostic Measures ofGoodness-of-Fit

Both the market analyst using regression and the userof AVM output must be aware of and understand howthe various key statistical measures used in regressionrelate to the reliability of results. These statistics fallinto two categories: overall measures that aid in theinterpretation of model performance and individualvariable measures that assist in the understanding ofhow well an individual variable performs in helping toestimate value, as well as keeping the standard errorterm to a minimum. Primary measures of goodness-of-fit for overall model performance are thecoefficient of determination (R2), standard error ofthe estimate (SEE), COV, and average percent error.

Goodness-of-fit measures for individual variables in amodel are produced by most MRA software packagesand include the coefficient of correlation (R), T-statis-tic, F-statistic, and beta coefficients. Each of thesemeasures will provide information about an individualvariable’s linearity or importance of contribution towardimproving predictive success, and relative importance,as variables are compared to each other.

(D’Agostino and Stephens 1986.)

When all the measures are used collectively, along withan understanding of data quality issues, those skilled indeveloping and using MRA can fully evaluate the cred-ibility of the AVM estimates. Appraisers asked to reviewAVM results must understand the role that goodness-of-fit statistics play in evaluating AVM results. The applicationof AVM results to a single property may be betterevaluated using historical market comparisons selectedfrom a subset of data. Appraisers asked to review AVMresults should review the Appraisal Standards Board’sUSPAP Standard and AO-18.

(Appraisal Foundation 2003, 46–56, 180–187; IAAO1990; D’Agostino and Stephens 1986.)

4.1.3 MRA Software, Options and TechniquesMRA is the most widely used method for calibratingmodels. As such, the availability of MRA software providesusers many choices. No one software package is deemedsuperior to another, as success using MRA is a combinationof modeling skills and software familiarity. Variations of a

selected MRA technique can be a decisive factor in selectingan MRA and statistical application package. Many MRAtechniques have been adopted over the years to helpregression take better advantage of its predictive powers.Stepwise, constrained, robust, ridge regression, and othersare acceptable techniques used to improve predictivesuccess. Many of the statistical software packages includevariable selection routines that aid the market analyst inselection of significant variables.

4.1.4 MRA Strengths1. Goodness-of-fit statistics—gives credence to

the validity of results.

2. Software availability—many regressionsoftware products are available.

3. Widely-accepted calibration method.

4. Broad education network—MRA is taught atmost colleges and universities around theworld.

5. Credible values—in the hands of a skilledmarket analyst, MRA is proven to produceresults that meet the test of model performance.

4.1.5 MRA Weaknesses1. Requires a high level of statistical

knowledge—market analysts must possesssignificant background in data analysis andstatistical methods.

2. Predictive accuracy is restrained by assumptions.

3. Requires data sets that meet the test ofsample size.

4. Interactive and nonlinear market trends aredifficult to measure without transforming data.

4.2 Calibrating Using Adaptive EstimationProcedure (AEP)

Adaptive Estimation Procedure (AEP) is a calibration tech-nique that was adapted to real estate value in the early 1980s.Also known as feedback, AEP is based on an engineeringconcept that relies on continual adjustment to coefficientsas the calibration engine passes, or tracks, back and forththrough the data until convergence, (minimum error isachieved) thus the feedback. For property valuation, thealgorithm tracks the sale price as a moving target. Itcompares property characteristics as variables that mea-sure the change in sale price, and calibrates a coefficient foreach variable. The coefficients are used to estimate valuethat is then compared to sale price. A running tally is kepton the error term as the process continues. Figure 1 depictsthe feedback loop.

AEP will make multiple passes through the sales fileconstantly adjusting coefficients before a final solutionis reached. Success using AEP is dependent upon the

13


market analyst’s ability to properly specify a model withcharacteristics that measure and evaluate local marketconditions. Market analysts using AEP have consider-able control over the variables used in the model and thecoefficient amounts. AEP uses whatever variables areintroduced into the model. No variable is excludedbecause of insignificance. As part of the specificationphase, the model can be pre-calibrated with starting,minimum, and maximum coefficients in order to help itconverge sooner, and to help ensure that rational coef-ficients will be produced. Setting the starting, minimum,and maximum coefficients is analogous to constrainingcoefficients used in constrained regression.

4.2.1 AEP Model StructureA hybrid model structure has the ability to directly deal withinteractive and nonlinear effects found in the market place.The structure closely resembles a cost model; however, thecalibrations give it the benefit of a direct market model. Theflexibility of the hybrid model built into AEP allows thequalitative variables to be calibrated in two different ways:as multiplicatives, that is XiB1 (rates), or binaries B1Xi.Deployment of a feedback model in an AVM format allowsfor flexibility without the added complexity of transforma-tions found with additive models.

4.2.2 Variable Control in AEPCalibration of individual variables in AEP differs signifi-cantly from the fitting of a straight line or curve in linear ornonlinear regression. Controlling for extremes in the coef-ficient amounts is a concern when using feedback. The useof smoothing and damping factors will help provide modelstability during the calibration phase. Smoothing is applied

to only the quantitative variables. Using an algorithm,smoothing keeps track of each variable’s exponentiallysmoothed mean (moving average) as a way of learning untila final solution is reached. Smoothing factors are used inconjunction with damping factors. The market analystprovides the settings for the smoothing factor. Additionally,damping factors control the amount of movement eachcoefficient (quantitative and qualitative) will have as eachnew case is introduced into the model while calibrating.Some feedback systems will dynamically adjust dampingand smoothing for optimized results. Locking or constrain-ing coefficient movement, forces residuals onto anothervariable. With so much control over the model, evensimilarly specified models may produce different finalanswers.

4.2.3 Results and Goodness-of-FitMeasures

Final results using AEP are measured first by the compari-son of how close the estimated price comes to the actualprice. Another measure, the reasonableness of coefficientamounts, is based on the skill and knowledge of the analystin pre-defining the model prior to calibration. AEP does notcare if a model uses square foot of living area at a price orthe window count at a price. If either can logically predictaccurate value estimates, AEP will generate a coefficientthat produces the lowest error term. Feedback understandsthat grouped patterns of property characteristics are thedeterminants of price and the individual characteristics donot necessarily produce marginal contributions to price.

The AEP is not reliant on statistical measures of themodel, or variable significance. Convergence occurs

AEP–Feedback Loop

1. Get a Sale Parcel

2. Predict Value

3. Compare to Sale Price

4. Adjust Coefficients

FIGURE 1.


14

when the average absolute error does not change appre-ciably from one iteration to another. Some softwareallows other criteria to be set by the user (e.g., maximumiterations, pre-defined absolute error). There is no sta-tistical measure that accounts for significance of avariable. Pseudo-R2 statistics can be generated after thefeedback model is complete. Output should include theaccounting for calibration of each variable by givinginformation on the number of observations, starting,minimum and maximum ranges, and the low, high, andfinal coefficient of the variable.

4.2.4 AEP Advantages1. Produces separate estimates for land and

improvements.

2. Based on reducing the absolute error term,not just minimizing the squared error term.

3. Outliers’ influence can be diminished duringvariable calibration cycle.

4. Requires fewer observations than regression.

5. Individual variable movement can be easilyconstrained.

6. Cost system attributes can be directly calibrated.

4.2.5 AEP Disadvantages1. Software availability is limited, and there is

no standardized algorithm.

2. Does not contain standard goodness-of-fitstatistics found in regression software.

3. Requires initial model be specified carefully.

As an alternative, some market analysts have turned to usingnonlinear regression software. Nonlinear regression supportsthe hybrid model and can calibrate interactive effects andcurves simultaneously like the AEP/Feedback routine.

(Ward and Steiner 1988; Gloudemans 1999, 196; Wooleryand Shea 1985; Carbone 1976.)

4.3 Artificial Neural NetworksThe most recent adaptation for use in calibrating real estatevaluation models is Artificial Neural Networks (ANN). Theconcept is borrowed from the biological sciences andfunctions of the human brain. The key element of the ANNparadigm is the novel structure of the information process-ing system. It is composed of a large number of highlyinterconnected processing elements that are analogous toneurons and are tied together with weighted connectionsthat are analogous to synapses. As the name implies, ANNcomes as close to producing artificial intelligence models asany calibration method. Like nonlinear regression andfeedback, neural networks can calibrate models that consistof both linear and nonlinear terms simultaneously. The userinputs each variable with assigned weights (coefficients).The software exposes the data using an algorithm in a

hidden layer where the weights are adjusted (calibrated) ina manner that reduces the squared error. This is an iterativeprocess much like those found with feedback and nonlinearregression. The final output results in a single estimate ofvalue with the exact formula remaining hidden from themarket analyst.

4.3.1 The Artificial NeuronThe basic unit of neural networks, the artificial neurons,simulates the four basic functions of natural neurons.Those functions are represented by inputs, the process-ing of inputs (summation), transfer (linear, sigmoid,sine, and so on), and outputs an answer. Artificialneurons are much simpler than the biological neuron;Figure 2 shows the basics of an artificial neuron.

Inputs to the network are represented by the math-ematical symbol x(n). Each of these inputs aremultiplied by a connection weight that is representedby w(n). In the simplest case, these products aresimply summed, fed through a transfer function togenerate a result, and then output.

Even though all artificial neural networks are con-structed from this basic building block, the fundamentalsmay vary in these building blocks.

4.3.2 Strengths of Neural Networks1. The ability of the neural network to “learn”

as it goes and to take new information andprocess as it has been trained.

2. Neural networks can recognize and matchcomplicated, vague, or incomplete patterns in data.

3. Options that provide analysts confidence aboutfuture use of neural network applications, suchas helping to improve data quality.

4. Studies completed indicate that the accuracyof neural networks is comparable to othercalibration methods found in the standard.

4.3.3 Weakness of Neural Networks1. The complexity of how the process actually

works in the hidden layer.

2. Lack of a definable model structure at theoutput stage makes explanation of value andsupport of the value more difficult.

3 . Requires considerable background indata analysis, data structure, andmathematical concepts.

4. Limited research links pertaining to use inreal property valuation.

5. Requires considerable investment incomputer power and software.


15


4.4 Time Series AnalysisTime series analyses are a family of techniques that canbe used to measure the cyclical movements, randomvariations, seasonal variations, and secular trends ob-served over a period of time. In property valuation, theseanalyses can be used to develop a multiplier or indexfactor to update existing appraised values or to adjustsales prices for individual properties to the valuation date.Since values can change at different rates in differentmarkets, separate factors should be tested for eachproperty type and market area.

Four methods used to develop time trend factors in theappraisal and assessment industries are: (1) value per-unitanalysis, (2) re-sales analysis, (3) sales/assessment ratiotrend analysis, and (4) inclusion of time variables in salescomparison models. These methods are summarized below(for a more detailed explanation and discussion, see MassAppraisal of Real Property (Gloudemans 1999, 263-270).

Value per-unit analyses track changes in sale price perunit (e.g., per square foot for residential properties or perunit for apartments) over time. The method is easilyunderstood and lends itself well to graphical representa-tion, as well as to statistical modeling to extract theaverage rate of change. A downside is that the methoddoes not account for the myriad of other value influ-ences, such as age and construction quality, that impactper-unit values.

Re-sales analysis uses repeat sales occurring over agiven time period. Price changes between sales areconverted to monthly rates and an average (or median)rate of change is extracted. As can be imagined, thelarger the number of repeat sales, the more reliable theestimated rate of change. The method can overestimaterates of change if repeat sales reflect substantial im-provements (or other alterations) made to the propertysince the first sale.

Sales/assessment ratio trend analysis involves trackingchanges in the ratio of sales prices to existing assess-ments made as of a common base date. Increases in theratios indicate inflation and vice versa. The ratio alsoprovides the index factor required to convert assessedvalue to a full value estimate. Like value per-unit analy-sis, the method lends itself well to graphical and statisticalanalysis. An advantage of the method is that assess-ments account for most value determinants and thuscan isolate time trends better than the value per-unitmethod. The method assumes that the assessmentsshare a common basis, and its reliability depends partlyon the accuracy or uniformity of the assessments.

Time variables can be included directly into AVMmodels to capture the rate of price change over theperiod of analysis. This is usually the most accurate ofthe various methods. However, model developers mustbe careful that time variables are properly specified sothat coefficients developed from the model reflect thedesired valuation date.

Once a time trend is established, it can be used to adjustvalues to any point within the sales period.

Trend factors can be extrapolated for a short periodbeyond the sales period, but this must be done withcaution and grows increasingly unreliable as the timeframe is lengthened. If more than several months areinvolved, the first three methods can be used to calibratethe trend (one would not ordinarily develop time adjust-ments through use of a modeling approach withoutrecalibrating the entire AVM model).

(The Appraisal Institute 2002, 291.)

4.5 Tax Assessed Value ModelTax assessed value models derive an estimate of valueby examining values attributed to properties by the localtaxing authorities. As a matter of local law and custom,

Sum

Inputs Weights

Transfer

Processing

Element

Output

Path

Y=f(I) Transfer

xnxn

I = ∑ w1

x1

Summation

wn

w0

w1

w2

x2

x1

x0

FIGURE 2.


16

the values reported by the taxing authorities often (butnot always) vary from the current market value in somereasonably predictable manner. For example, somejurisdictions require the taxing authority to report theassessed value at 25 percent of the estimated marketvalue. Some jurisdictions may not have reappraised in along time, so values lag far behind the current market.Also, some jurisdictions report multiple values: as-sessed, appraised, and market values. By examininglocal laws and customs with respect to how values aredetermined, as well as applicable time trends (see Sec-tion 4.4) and information reported in local or state-levelratio studies (see Section 8.4), it may be possible todevelop adjustment factors to apply to values reportedby taxing authorities in order to approximate currentmarket values.

The reliability of a tax assessed model will depend on theuniformity of appraisals to which the adjustment factorsare applied, as well as the accuracy of the adjustmentfactors themselves, which can vary with how currentthe assessments are, and the reliability of the ratio studiesor other information on which they are based. Extremecaution must be exercised when local assessment uni-formity is poor, because factoring an unreliable assessedvalue will only result in an unreliable market value. On theother hand, local assessments that meet IAAO standardscan provide a sound basis for market values estimation.

4.6 Calibration SummaryThe various methods and procedures used to calibratethe AVM are the engines that drive accuracy andcredibility of the estimate made. By itself, no onecalibration method is better than another. Data integrityand the skill level of the analyst define the accuracy ofone calibration technique as compared to others. Usersof AVM products must be aware of the interdependencebetween skills and technologies of calibration whendeciding how well the AVM will perform.

The use of MRA has been the longstanding choice forcalibration and has a proven track record. Feedback,nonlinear regression, and neural networks are emergingtechnologies that require different levels of skill andknowledge concerning modeling real property values.Understanding calibration in relation to this standardencourages the AVM market analysts and clients tounderstand that AVM development is not a black boxprocess; instead, it is based on well-defined conceptssurrounding the appraisal process. Details for learningand understanding the skills and technical aspects ofcalibration are found in the references throughout thissection of the standard.

AVM clients must understand that developers of AVMproducts are not limited to using a single method ofcalibration. Product market analysts often base their valueestimates on multiple technologies. Included in these tech-

nologies are simple sales listings of automated sales com-parison selections, with adjustments derived from themodeling process. Appraisers asked to use or review anAVM should read Advisory Opinion 18—published as partof USPAP Standard 6 by the Appraisal Foundation (2003,46–56, 180–187).

5. RESIDENTIAL AVMSThe residential property class has the longest history ofbeing valued by AVMs. Residential property includesdetached single-family homes, condominiums,townhouses, and zero-lot-line property. Other propertytypes included in the residential class are properties withfour units or less. Traditional methods of valuing theseproperties are cost approach and direct sales compari-son. Both methods have been automated and areconsidered a part of the AVM category of methods andtechniques available.

5.1 Detached Single-FamilyWhen adequate sales data is available, the direct salescomparison approach is the preferred method of valuingresidential property. The approach may take two forms:direct market models and comparable sales.

Direct market models developed from sales analyses usevarious model structures, with coefficients derived viaa mathematical calibration method. The comparablesales method is a two-part method in which comparablesales are found and then adjusted to the subject prop-erty.

Some AVMs combine the strengths of direct marketmodels and comparable sales models, to the point wherecomparable sales model coefficients are derived fromdirect market model analysis.

Cost models, like sales comparison models, have astrong history of reliability and credibility for valuingresidential property. However, the origin and accuracyof coefficients are unknown to most users and may notreflect the actual market.

5.1.1 Cost ModelsThe cost approach works best when applied to newerproperties that do not exhibit a great deal of measurabledepreciation, and where the land value can be reason-ably estimated from recent land sales. Cost models areanchored in tables developed by studying local buildingcost data. In the AVM format, the tables are convertedto a formula and applied by simply entering basicbuilding (improvement) information. Such models areused for deriving the Replacement Cost New (RCN).The initial cost coefficients supplied with a cost modelrepresent the supply side of the residential market.These RCN estimates need further calibration for actualproperty condition (depreciation), location (macro and

17


micro), and a supportable estimate of vacant land value,in order to arrive at market value. These items representthe demand side of the market. A strength of the costmodel is that it can be applied to any improvementregardless of size, quality, age, condition, or style. Theaccuracy and credibility of the cost model is tied to theanalyst’s ability to calibrate depreciation, location, andland value.

5.1.2 Comparable Sales ModelsKnowing the sale price of a property with attributessimilar to the subject property is a concept thatconsumers can easily understand. This approachprovides the theoretical basis for the Sales Compari-son Model using comparable sales. Sales comparisonof residential property has been accepted by realestate consumers and the courts for many decades;however, this method does have limitations in theautomated world. It essentially requires two models.The first one is a comparable selection model. ManyAVMs rely on identification and summarization of allrecent sales within a specified radius of the subject.The advantage of this model is that all recent saleswith close proximity to the subject are considered.This method may work well in homogeneous areaswith a high sales volume. If the comparables havesignificant attribute differences, the confidence ofthe adjustments being made also begins to suffer.For quality comparables, an AVM routine may con-sider using a weighted selection model (e.g., regressioncoefficients, Minkowski or Euclidian metrics). An-other choice would be cluster analysis.

All of these methods can select comparables based onattribute comparisons that pick the comparables mostsimilar to the subject, based on defined parameters.These methods are not limited to selecting only threesales, as has been the tradition. Once the bestcomparables are selected, they must be adjusted forattributes that are dissimilar to the subject. Howthese adjustments are developed has much to do withhow accurate and reliable the sales comparisonestimate will be. Mathematically, the adjustmentscan be derived from just two sales; one sale pos-sesses the attribute, while the other does not. Thedifference in sale price measures the value of themissing attribute. Sales comparison methods thatrely on direct market models that use quantitativemethods for deriving the adjustments, are morestable and reliable than simple match pair analysis.

In its formatted form, the comparable sales approachshould display how each attribute adjustment in theAVM contributes to the overall value estimate. Users ofAVMs are cautioned that matched pairs analysis is nota statistical calibration method. Any comparable salesapproach claiming to be an AVM as defined in this

standard must meet the criteria of being supported by anautomated market analysis process.

5.1.3 Direct Market ModelsThe basic premise of direct market models (also termedhedonic models) is that the price of a marketed good isrelated to its characteristics, or the services it provides.For example, the price of a home reflects the character-istics of that home (e.g., size, construction quality, style,location). Therefore, we can value the individual at-tributes of a home by looking at the prices people arewilling to pay for them. Direct market models lendthemselves well to the calibration methods and tech-niques discussed in Sections 2–4 of this standard. If avalue-determining attribute can be captured in a data-base, then the model can calibrate a coefficient thatmeasures its contribution to the total value estimate. Prop-erly designed direct market models will produce AVMscapable of very accurate and credible value estimates.

All three model structures introduced in Sections 2and 3 are well suited to the valuation of single-familyresidences. Additive models have been the traditionalworkhorse and work very well in most cases. Multi-plicative models carry certain advantages discussedearlier and can also be effectively adopted. Becausethey accommodate dollar and percentage adjustments,hybrid models provide the most flexibility. Where anadditive model will add the same lump sum amount toall property having air conditioning, multiplicative andhybrid models will attribute different amounts de-pending on the style, quality, and location of theproperty. Both model structures also lend themselveswell to the valuation of spatially dispersed or highlyheterogeneous residences.

5.2 Attached Residential Property(Condominiums, Townhouses,Zero-Lot-Lines)

Structures built on an individually plotted lot designedfor only one family to occupy, are termed “detachedsingle family residences” and make up the majority ofresidential property in most communities. Zoning andother spatial changes in a community dictate the densityof residential land use. Other methods of dividing land,besides using land-based boundaries, lead to other typesof residential use and ownership. Structures wheremultiple living units are all joined together take ondifferent forms of ownership depending on how the titleis legally conveyed in the market place. These structuresare commonly referred to as “attached residential units.”A ten-story building with five units per floor could be aninvestment property with each unit rented. Propertydivided into air lots is known as condominiums. Anotherdivision of ownership rights is by time, where each day,week, or month represents units of ownership. Struc-tures where the ownership is divided vertically are


18

known as townhouses, row houses or zero-lot-lines,depending on geographic location throughout the world.All of these uses are residential in nature.

Valuing these various residential properties is somewhatsimilar to valuing detached single-family structures. Allof the same principles apply and all can be modeled andvalued using an AVM. In fact, because these propertiesexhibit a high degree of homogeneity compared to thedetached single-family population, sales-based AVMscan produce values that are extremely reliable andaccurate. The cost approach can also work well, insome cases, if adjusted to the market, but it is notappropriate for valuating condominium units becausedepreciated replacement cost will not properly reflectresale values. Data requirements for attached residenceswill not be the same as with detached residential prop-erties. For example, floor level can be an important valuedeterminant for condominiums, while lot size and yardimprovements are irrelevant.

5.3 Two- to Four-Family Residential PropertyPart of the residential housing market consists of struc-tures built for the purpose of housing more than onefamily. Improvements designed to accommodate two,three, and four families within their own separate livingareas are often referred to as small income-producingproperties. A common theme among these propertytypes is that the owner of the property may reside in oneof the units. This concept, however, is not a requirementfor classifying these structures in the market. Two-unitproperties are more likely to be owner-occupied thanfour-unit properties. The concept to be recognized hereis how such properties are treated in the marketplace,because that impacts their price and ultimately the valuegenerated by any AVM. The ability to model the sellingprice of these small-income properties is reliant on whatspecific data is available, relating to number of units, age,condition, location and gross income. The motivation ofbuyers shifts when consideration is given to otherproperty attributes that relate to producing rental incomeand not just owner occupancy. Direct market models,comparable sales models, and cost models are acceptablemethods for valuing these small income-producing proper-ties. With their income-producing potential, the incomeapproach is also a model to be considered. With an adequatesample of gross income values for comparison to saleprice, a model of GI * GIM will yield credible results whereGI = Gross Income and GIM = Gross Income Multiplier(sale price/gross income). Some AVMs may even be set upto predict GI and the GIM. Each of these indicators can varywith size, age, location, style, and condition of a property.

5.4 Manufactured HousingA manufactured home is a residential structure built in afactory. Construction standards for manufactured hous-ing are controlled and monitored by the Department of

Housing and Urban Development in the United States(HUD), and by the Canada Mortgage and HousingCorporation (CMHC) in Canada. While many manufac-tured homes are built with the same materials as site-builthomes, the factory-controlled engineering process helpscontrol cost and quality. The house can be financed aspersonal or real property on leased land, in a manufac-tured home community, or on a privately owned site.Buyers who desire to acquire land in conjunction withthe home can finance the land and home together.Market conditions and trends will indicate how themanufactured homes compete in the market place. Insome communities, zoning only allows manufacturedhomes in certain areas, confining the market area fromwhich comparables can be derived. Once market con-ditions for a manufactured home are known, it can bemodeled just like any other property type. Consistencyis important when using an AVM for manufacturedhomes. Some manufactured homes are strictly treated inthe market as mobile homes (i.e., personal property). AnAVM developed to value manufactured homes as realproperty would give a false value in the case where thehome was personal property, and vice versa. AVMsdeveloped to value manufactured personal propertyhomes cannot be used for homes classified as realproperty. Some manufactured homes compete in themarket place with site-built homes. Where this is thecase, it is possible that an AVM designed to valuedetached single-family structures will produce credibleresults, although the model should include a variable (orvariables) to capture any differences between otherwisecomparable manufactured and site-built homes.

5.5 Time Series Models for Residential PropertyIndexed models relate to time-series analysis (see Sec-tion 4.4 on Time Series Analysis) as described earlier.Use of these models represents a common method ofdelivering quick automated value estimates. These mod-els simply measure the average change in value over timeand factor the value forward from a benchmark starting-place, such as the average value in a census block ormarket area. The accuracy of indexed models is incon-sistent and less reliable than fully specified models.These models work best in areas of homogeneity wherethe range of value is close to the average value.

Indexing is a common method used to update cost tablesto reflect current cost. As with market models, abenchmark in time is required as a starting point. Costcoefficients are then updated, using a single index factorrepresenting the measurable change since the originalcost coefficients were generated. One current methodof indexing is to use an economic indicator such as theconsumer price index (CPI). In the cost approach,indexed models have no way of adjusting values at themicro level for location and other market influences thatimpact value. Time adjustments may be developed from

19


the analysis of known sale prices within a geographicarea, such as a neighborhood or postal code, and over aspecified time reference.

Users and consumers of index models must understandhow the index factor is created and how the accuracy ofthe original value was derived before giving a lot ofcredibility to an AVM using an indexed model.

5.6 Summary and Conclusions for UsingResidential AVMs

AVM developers (or users) must understand the in-tended use of the residential property. The residentialhousing market is diverse. AVMs lend themselves toestimating the value of residential property. However,each class of residential property has some uniquecircumstances that will influence how well the AVM canperform when estimating the value. When the unique-ness is captured as part of the data used to develop theAVM, the chances of the value estimate being accurateand credible increase greatly. When unique characteris-tics are ignored, they are not measured in the market andthe error term of the values produced will increase,destroying confidence in the AVM’s ability to estimateaccurate and credible values.

The overall ability of the AVM to accurately estimatevalue can be evaluated using the quality assurancemeasures found in Section 8 of this standard. If theassurance standards are being achieved, then the validityof the AVM is known, and the market analyst and userscan understand what degree of confidence to expectfrom the ensuing value estimates.

6. COMMERCIAL AND INDUSTRIAL AVMSCommercial and industrial properties, including apart-ments and multifamily residences with greater than four(4) units, are usually income-producing properties ac-quired for their ability to generate income. As a result,commercial and industrial properties are best valuedusing an Income Approach where adequate income dataare available or the sales comparison approach whereadequate sales are available. However, a solid CostApproach is needed where sales and/or income data areinsufficient to calibrate an appropriately structured model.Also, care must be taken in developing and applyingincome valuations, to appraise only the real property andnot the business, and to value based on typical manage-ment, not on the present management.

Commercial and industrial properties provide their uniqueAVM challenges. First, in some markets there arerelatively few sales of commercial and industrial prop-erties. This creates problems with land valuation for thecost approach, development of comparable sales orstatistical models, and for developing capitalization ratesand multipliers for the income approach. The market

analyst may have the additional problem of needing toprovide separate land and building values that mostincome models are not designed to deliver.

Location for commercial and industrial properties canrange from relatively little effect to extremely important.

Finally, special purpose properties and limited marketproperties, such as theme parks and casinos, are gener-ally included with commercial and industrial properties.These properties tend to be unique and, as a result, aredifficult to categorize and value.

6.1 Commercial and Industrial ModelSpecification

Valuation of commercial and industrial properties re-quires market and income or cost data. Income andmarket data are preferred. Cost data is needed whereinsufficient sales and income data are available. Com-mercial and industrial sales comparison models, likeresidential models, require data on use, location, andphysical characteristics.

6.1. Property UseThe property use is extremely important as a comparisoncharacteristic. It is necessary to determine the generalcategory of property use. The property use does not needto distinguish detailed specific uses, such as shop versusliquor store or gift shop. Use of broad categories willincrease the number of properties for which informationcan be captured, analyzed, and compared.

6.1.2 LocationAs with residential properties, location can be includedeither through the use of neighborhoods or market areaswith binary (dummy) variables or categorical variableswith percentage adjustments, LVRSA, or as a distancevariable to Value Influence Centers (VICs), such as thecentral business district. For commercial and industrialproperties, location analysis relates largely to identifyingzones or groups of properties subject to similar influ-ences. Proximity to VICs is important in commercial andindustrial valuation, but a lack of commercial and indus-trial sales may make location of the VICs, as well asmeasuring their effect, difficult.

The importance of a location adjustment will alsovary considerably with the property use. For ex-ample, while the value of a service station generallydepends on location on a major street, such a variablemay not be needed if all service stations throughoutthe jurisdiction, or area, enjoy such locations. Otheruses, such as hotels, may be highly dependent onlocation, such as being on a beach area or near aconvention center. Finally, what is considered anuisance for residential properties, such as a railroadtrack or heavy traffic pattern, could be an importantamenity for commercial and industrial properties.


20

The most common method of modeling location isthrough the delineation of economic areas or neighbor-hoods. Central Business Districts, cities/towns, andother areas of significant deviation from the norm, areused to identify economic areas or neighborhoods.When using economic areas or neighborhoods forlocation adjustment, care must be taken not to create toomany as this may result in too few sales or insufficientincome data for analysis and modeling.

LVRSA using GIS or manual grids is also used for devel-oping location adjustments where sufficient data is available.

6.1.3 Physical Characteristics andSite Influences

Commercial and industrial properties require a numberof physical characteristics for comparison and model-ing. These may be quantitative or qualitative variables.

The most significant quantitative characteristic is build-ing area. Different building areas may be used for cost,sales comparison, and income approaches. Areas aredifferentiated by type, such as basement, ground floor,and upper floors, for cost valuation; whereas incomemodels generally use net rentable areas, differentiated byuse, such as retail, office, etc. Sales comparison modelsalso benefit from use differentiations, although eithergross or rentable areas can be used. Some sales com-parison and income AVMs utilize other units ofcomparison, such as units for apartment buildings,rooms for hotels, and spaces for parking garages. Otherkey quantitative variables are the year built and effectiveage or condition, which are used to capture accrueddepreciation and Remaining Economic Life (REL). Ef-fective age (EA) or REL is a critical factor of comparisonfor cost, market, and income modeling. The EA or RELis also a key variable in determining the relationshipbetween income and value because it establishes the timeremaining for the income stream.

Other significant quantitative and qualitative variablesare similar to those used for residential AVMs. Suchexamples include building quality and lot size. Whileothers, such as traffic patterns or ceiling height, may beimportant to specific property uses or occupancies.

6.1.4 Income DataAn income value is essentially a calculation of the presentworth of the future benefits to an income stream. It isused to estimate the market value of a property based onwhat an investor would pay for the property. Incomedata includes revenues, expenses, net income, andcapitalization rates or income multipliers, which are thenused to develop a projection of an income stream toestimate the market value.

The income value is generally estimated by either capi-talizing the Net Operating Income (NOI) or developing

a multiplier for the potential or effective gross income.The capitalization rate can be developed as an OverallCapitalization Rate (OCR) from the market place bycomparing the estimated NOI against sales prices,where available. Sales and gross income data can beused to develop a GIM from the market place. GIMs canbe accurate, require less data, and eliminate the need forexpense analysis. They can be developed from potentialor effective gross income as long as the data is collected onthe same basis. While GIMs may be easier to develop,overall rates and NOIs may more accurately and directlyreflect the value of the income stream critical to investors.

Due to the sensitivity of income data, the widely varyingmanner in which it is kept, and the differences ininformation maintained for differing property types,income data is difficult to ascertain. Creating differentreporting forms for different property types makes theforms easier to use and understand, thereby increasingthe likelihood that more forms will be completed andreturned. Breaking income and expenses into genericcategories also facilitates reporting. However, creatingtoo many categories may only complicate the form andminimize the number of completed returns while notnecessarily contributing to a more accurate net incomecalculation. Minimizing the detail collected, includingavoiding tracking information about individual tenants,serves to make the data more likely to be completed andeasier to maintain.

The pool of sales and income data can be expanded byusing multiple years of data and making any indicatedtime adjustments. However, if the income and sales dataare from the same time period, neither needs to beadjusted for time for the purposes of developing capitali-zation rates and income multipliers. In addition, tradepublications and local banks may serve as sources ofinformation to build capitalization rates and multipliers.

6.2 Development of the Model(s)Commercial and industrial properties can be valued bysales comparison, income, and cost AVMs. Becausethere are fewer commercial and industrial sales, it isoften difficult to develop comparable sales and statisticalmarket models for commercial and industrial properties.However, a number of income-approach models may bedeveloped using sales to develop capitalization rates, andGIMs using gross incomes and expenses derived fromthe local market or industry-specific publications whenlocal data are insufficient. The cost approach, whilegenerally the least desirable, is still necessary for prop-erty types that have insignificant sales and insufficientrevenue or expense information.

Income models may be developed using stratification orglobal methods. Stratification requires grouping com-mercial and industrial sales by factors that affect therelationship between income and value. This is accom-

21


plished by groupings based on use or occupancy, age orcondition, and location. As with any valuation approach,the more strata you create, the fewer data in each strataare available for analysis. The use of global methods,such as MRA, can be used to overcome the limited datain many strata by combining selected property types(such as all retail-related properties) into a single modeland using binary variables to differentiate the specificuses or occupancies (such as general retail, restaurant,or convenience mart).

Industrial properties may be modeled in the same man-ner as commercial properties, but there are even fewerindustrial sales than commercial sales. Often ware-houses and light industrial properties can be combinedinto a single model to increase sample sizes.

6.2.1 Cost ModelsWhile commercial and industrial cost models are similar toresidential cost models, they typically comprise differentstructural components. The commercial and industrial costmodel requires a number of extra features or miscellaneousitems. Cost models are most appropriate for commercial,industrial, and special purpose properties where there isinsufficient sales and income information.

6.2.2 Sales Comparison ModelsIt is often difficult to get sufficient qualified sales to developcommercial and industrial comparable sales and statisticalmodels. However, where sufficient sales can be found,direct market models can be developed using variables forlocation, size, construction quality, age or condition, landsize or frontage, and relevant amenities or nuisances.Additive, multiplicative, and hybrid models can all be used;yet proper model specification is critical.

6.2.3 Income ModelsThe income approach can be used to develop commer-cial and industrial AVMs. Because these properties arefrequently sold based on their income streams, theincome approach can be the most desirable. The twomost popular approaches are direct capitalization andGIMs. Discounted cash flow (DCF) analysis can also beused; however, the data requirements for developingyield capitalization estimates from DCF analysis makethe method more challenging than direct capitalization.Also, a number of the assumptions required for DCFanalysis, including anticipated yield, holding period, andvalue at the end of that period, can be difficult to derivefrom the market and, therefore, may be subjective.

6.2.3.1 Modeling Gross IncomeGross incomes may be analyzed from local market surveysor questionnaires or they may be obtained from industrypublications. Typically the gross rent per unit (e.g., squarefeet/square meters, rental unit, or room rate) is the depen-dent variable in the model. Gross income models are

ordinarily easier to develop than net income models becausethe data is easier to obtain and less subject to manipulation.Gross income models can be developed for either potentialor effective gross incomes. The independent variables arethose that affect the expected gross income, including:location, age or condition, amenities and nuisances, etc.Where data is limited, to develop separate models for eachuse or occupancy, a single model may be developed bydetermining a reference use or occupancy group and usingbinary or categorical variables for the other use or occu-pancy groups.

6.2.3.2 Vacancy and Collection LossesVacancy and collection losses are deducted from thePotential Gross Income (PGI) to account for typicallosses due to vacancy and bad debts based on localmarket conditions. The vacancy and collection lossusually varies by property use and is expressed as apercentage of the annual PGI. The percentage may bedetermined by a market analysis of PGIs compared withactual income, or from information supplied by locallenders and industry trade publications.

6.2.3.3 Modeling ExpensesExpense data may be obtained from the same sources asthe revenue data. Expense ratios can be developed byeither stratification or a modeling approach. The ex-pense ratio is the dependent variable, and the independentvariables are similar to (but typically fewer than) thoseused to determine the gross income per unit. Like grossincome models, a single expense ratio model may bedeveloped, where insufficient data are available formultiple models, by determining a reference use oroccupancy group and creating binary variables for theother use or occupancy groups.

6.2.3.4 Direct CapitalizationDirect capitalization involves developing an overall rate(OAR) directly from the market place. The OAR is thenused with the estimated net income to estimate the valueby income capitalization. Like expense ratios, capitaliza-tion rates can be developed using either stratification ora modeling approach. The advantage of OARs is thatthey use the NOI that includes both gross incomes andexpenses, and thus may specifically reflect a typicalinvestor analysis of commercial properties. The depen-dent variable in developing a direct capitalization rate isthe indicated OAR (estimated net income divided by thesale price). In developing an OAR model, a single modelcan be developed by determining a reference use oroccupancy group, and creating binary or categoricalvariables for the other uses or occupancy groups. As inthe revenue and expense models, this permits more datato be used in the model. In addition to variables forlocation, age or condition, and amenities and nuisances,the OAR model should include an adjustment for at-


22

tributes that affect the recapture portion of the OAR(such as land/building value ratios, REL estimates, andexpense ratios).

6.2.3.5 Gross Income MultiplierGIM models involve developing multipliers directly fromthe market place for either the potential gross income or theeffective gross income, depending on the data collected.Effective Gross Income Multiplier models are generallymore stable. GIMs have the advantage of not requiringexpense data that may be missing, unreliable, difficult tointerpret, or incomplete. The GIM is the dependent variablewhile the independent variables are typically the same asthose previously described for an OAR model. However, itis important to ensure that variables related to differencesin expense ratios are included because gross incomes areunadjusted for expenses.

6.2.3.6 Property TaxesCare should be taken to treat property taxes consistentlyin the development and application of AVMs. Propertytaxes may be included as an expense or as a componentof the OAR.

6.3 Quality AssuranceCommercial and industrial quality assurance is particu-larly critical due to the limited amount of sales andincome data available for analysis and modeling. Com-mercial and industrial quality assurance is accomplishedin much the same manner as with any other type ofproperty. Valuation research and appraisal proceduresare subject to review, and the values tested and statisti-cally analyzed for accuracy and consistency.

In addition, quality assurance must be extended to theincome data collected. Estimated gross incomes, ex-pense ratios, OARs, and GIMs should all be reviewedfor consistency. Gross income and expense data shouldbe compared with like properties to identify outliers thatmay need to be removed from the modeling processunless the data can be corrected.

7. LAND MODELSIf ample sales are available, vacant land is generallybest valued using a sales comparison approach. Themost significant exceptions to this are leased land andrural/agricultural land that are usually valued using theincome approach.

Land provides a set of unique problems for AVMs. Landis highly speculative and there frequently are relativelyfew sales for analysis and modeling (see Section 2.3.3on Data Management and Quality Analysis).

Land values are highly affected by location. This is also oneof the reasons why land values appear to be more specu-lative. Other factors affecting land values include Federal,

state, and local regulations affecting development and whatstage the neighborhood is in its life cycle. Developed landwill command a significant premium over underdevelopedland, especially when there is no guarantee that the pur-chaser or potential developer will be successful. In addition,neighborhoods evolve from growth to stability, to decline,and potentially to being a land-driven market where theimprovements have no value.

GIS is extremely valuable as an aid in establishing the effectof location on land. Where a GIS is not available, neighbor-hoods can be developed based on appraisal judgment, orgrids can be developed and x, y coordinates manuallyderived from the grids to better handle location.

Land that is significantly distant from urban areas maybe best valued based on its income potential.

7.1 Land Valuation Model SpecificationMarket land valuation modeling requires data on use,location, and physical characteristics. Land models, likeimproved models, require qualitative and quantitativevariables, as well as data transformations.

7.1.1 Property UseThe analyst must estimate the property use of a parcel ofland for any AVM. This will serve to determine how itshould be appraised as well as provide a key variable forcomparison and to determine what sales are best suited forbuilding the model. Although many states and provincesprovide use codes for reporting, currently there are nogenerally accepted standards for classifying land uses. TheAmerican Planning Association (APA) has recently pro-vided an update on their Web site of the 1965 Standard LandUse Coding Manual (APA 2003). However, the APA is notan appraisal organization and its solution contains multipledimensions—whereas appraisers generally focus on thecurrent use and the highest and best use.

7.1.2 LocationLocation and parcel size are arguably the most importantpieces of land data. The most common method of modelinglocation is through the delineation of economic (or submarket)areas or neighborhoods. More recently, variations of LVRSAhave been developed to determine location adjustments bothwith and without delineating economic areas.

Appraisers, using maps and their judgment (based onknowledge of market conditions), generally decide neigh-borhood or submarket boundaries. All parcels in aneighborhood or submarket receive the same locationadjustment. There are two factors to be aware of whenusing this approach. First, boundaries may be drawn tocoincide with major streets, natural barriers, and/orpolitical subdivision boundaries. And, secondly, themarket analyst should be aware that location adjust-ments can change abruptly from one submarket orneighborhood to another.

23


LVRSA is another method of developing location adjust-ments. LVRSA uses a geographic grid to display valueresiduals or sales ratios based on values derived from amodel lacking a location variable, to develop factors thatquantify the relative locational advantage or disadvan-tage of the property. This process may include identifyingpositive and negative VICs. Distance variables from allof the VICs are computed for each parcel. If VICs areused, the distance variables are then included in themodel to calculate the location adjustment for eachparcel. The location adjustments determined in thismanner may be developed for cost, sales comparison,and income AVM models. Due to the method LVRSAuses to develop the location adjustment, it will includeanything that is not accounted for elsewhere in itsestimate of the location adjustment.

Geographic grids for LVRSAs are best obtained froma GIS. However, where one does not exist, a geo-graphic grid can be manually developed by usingmaps and arbitrary grids, such as every 100 feet. Thex, y coordinates can then be determined for eachparcel and entered into the database. Although not asaccurate and effective as a GIS, this approach can beused where one does not exist or is not yet availableto the market analyst.

When using neighborhoods or submarkets for locationadjustment, care must be taken not to create too manyneighborhoods or submarkets; because this may resultin too few sales for effective analysis and modeling.Central Business Districts, cities/towns, natural fea-tures, and major streets can be used to defineneighborhood boundaries. Because the single propertyappraiser generally values only a single, or few proper-ties, and the AVM market analyst must value manyparcels and sometimes deal with adjacent parcel reviewby the public, the AVM market analyst might prefer touse blocks, subdivisions, or neighborhoods for locationadjustments so that adjacent and nearby parcels receivethe same adjustment.

7.1.3 Physical Characteristics and SiteInfluences

In addition to land use and location variables, AVMsrequire a number of physical characteristics and siteinfluences for comparison and modeling. These may bequantitative or qualitative variables.

The most significant quantitative characteristic is land size.Land size is determined by the number of land units by typesuch as lot, site, front feet or meter, square feet or meter,acre, hectare, etc. Therefore, it is usually necessary todevelop some form of land size adjustments to reflect thechanging rate per unit based on the total parcel size.

Most of the other important characteristics and influ-ences are qualitative. These include topography, site

amenities (such as government services), property ac-cess, water and sewer, proximity to negative influences(like railroads or treatment plants), and proximity topositive influences (like view, golf courses, water front-age, or recreational areas). However, keep in mind thata negative influence under one condition might beconsidered positive in another situation. An example mightbe high traffic volume that could be positive for commercialproperties and negative for residential properties.

7.2 Land Data CollectionSales and income data for land are collected, verified,and maintained in the same manner as improvedparcels. Maps and aerial photographs are used tosupplement field reviews to effectively collect, main-tain, and review land data.

Land use/soil productivity data for income modeling ofagricultural property may be obtained from Federal andstate/province agricultural agencies, universities, andagricultural cooperatives and associations. When insuf-ficient arms length sales are available, data to developcapitalization rates for agricultural properties may beobtained from farm lenders such as the Federal LandBank and Farm Credit Bank, as well as local lenders.

7.3 Development of the Model(s)Sales comparison is the primary approach for estimatingthe market value of land. The valuation of land by salescomparison shares many of the same analyses andmodeling processes with improved valuation models.The dependent variable in a sales comparison modelshould be sales price or sales price per unit. Forexample, if land sales in an area are based on square feetof land area, then the dependent variable should be saleprice per square foot. Typical independent variablesinclude property use, zoning, size, or location; sitecharacteristics including physical characteristics; ameni-ties (positive influences); and negative influences.

For leased land, and agricultural and rural properties, whereinsufficient sales are available, a capitalized income streamis commonly used to estimate the market value. Incomeland appraisal relies on capitalized income analysis.

7.3.1 Land Valuation Modeling by SalesComparison

Land values may be modeled separately from improvedvalues, or vacant and improved property may be mod-eled in a single combined valuation model (Guerin 2000).The primary benefit of a combined model is that bothvacant and improved sales are used, which significantlyincreases the sales sample size for analysis and model-ing. When developing a combined model, a binaryvariable should be used to separate vacant and improvedsales. In addition, separate time and size adjustmentsshould be tested for vacant and improved sales.


24

7.3.2 Land Valuation Modeling by IncomeIncome data can be used to value rented or leased land.Income capitalization for land follows the same generalprinciples as commercial and industrial properties.

8. AUTOMATED VALUATION MODELTESTING AND QUALITY ASSURANCE

AVM testing and quality assurance is necessary todetermine the applicability of the model and/or the needfor further specification. The process of developing anddeploying an automated valuation model must includesafeguards to insure the accuracy of data used and theintegrity of results produced. Those safeguards aresimilar in kind and effect to those employed in evaluatingthe performance of any mass appraisal project.

8.1 Data Quality AssuranceAll data used in model specification and calibration mustpass the following screening tests:

1. Data must be sufficient to produce reasonablepredictive models with regard to the propertycharacteristics utilized in model calibration andimplementation. As a general rule, the number ofsales should be at least five times (fifteen times isdesirable) the number of independent variables(Gloudemans 1999, 127).

2. Sales data must reflect, to the maximumextent possible, the conditions requisite tomarket value transactions.

3. Subjective data must be consistent across thepopulation of properties to be valued usingthe model. Examples would include quality,physical condition, and effective age.

4. Accurate property characteristic data isessential to model quality. If the data were tobe verified through a field audit, it should befound to be correct 95 percent of the time.

Data quality assurance should measure the quality andquantity of data, as well as provide a means of evaluatingthe application of the developed AVM formula to aspecific population of properties. The product of thatevaluation may include the acceptable ranges of specificproperty characteristics and ranges of estimated marketvalues to which the model can be applied.

In addition to the quality assurance statistics discussedbelow, it is good practice to provide the user with a measureor index of the relative confidence that can be placed inindividual value estimates, especially at the extremes of thedata ranges. Using stratified ratio studies to examine theextreme low and high ends of various property character-istics in the modeling and holdout data sets, the market

analyst will be able to determine the applicability of themodel at these extremes.

The market analyst should decline to provide an estimateat those points where the value estimates becomeunreliable due to the data falling outside of acceptableparameters (see Section 8.4 on Sales Ratio Analysis).

8.2 Data RepresentativenessBecause AVMs use a relatively small sampling of prop-erties from which inferences about the total populationof properties are drawn, care must be taken to ensurethat the sample adequately represents the total popula-tion of properties to be valued. In many kinds ofstatistical studies, samples are selected randomly fromthe population to ensure representativeness. Becausesales do not represent true random samples, extra caremust be taken to ensure representativeness. A sample isconsidered representative when the distribution of val-ues of properties in the sample reflects the distributionof values in the population. Because the distribution ofvalues in the population cannot be directly ascertainedand appraisal accuracy may vary from property toproperty (depending on property type and characteris-tics), representativeness can be achieved by selecting asample that adequately reflects salient value-relatedproperty characteristics. A property should be includedin a sample based on characteristics of the property andnot actions or characteristics of the owner.

This same degree of care should be taken in selectingsales samples used to test the quality of the AVMonce it is developed

(IAAO 1999, 12.)

8.3 Model DiagnosticsThe specific diagnostic tools available to market analystsand users of automated valuation models will vary withthe model methodology employed. Multiple regressionanalysis provides the market analyst and user with awide range of diagnostic statistics that may not beavailable with other calibration methodologies. In anyevent, the market analyst must make effective use of thediagnostic tools available during model calibration and beprepared to explain their use and significance to endusers.

Standards do not exist for goodness-of-fit statistics(such as the coefficient of determination) or measuresof individual variable significance (such as the T-statis-tic). Nonetheless, the market analyst should be able toexplain how those statistics were used and how theyrelate to the predictive quality of a specific model inrelation to the sales data available for calibration.

8.4 Sales Ratio AnalysisSales ratio analysis is a type of statistical study basedon comparisons between an estimated value and

25


market value as indicated by sales prices. For AVMuse, the numerator would be the estimated valuegenerated from the model, while the denominatorwould be the sale price. The ratios thus calculated aresubjected to statistical analysis to determine centraltendency (level), and vertical (value related) andhorizontal uniformity or variation. Central tendencystatistics provide information about the overall ortypical level in relation to market value that would beachieved given the results of the model. Variabilitystatistics provide information about the degree towhich model-determined values for individual prop-erties are similar with respect to market value.

Sales based ratio studies are among the most objectivemethods for testing the performance and quality of anymass appraisal system. Much of the information in thissection has been reprinted from the Standard on RatioStudies (IAAO 1999).

8.4.1 Measures of Appraisal LevelStatistically, measures of central tendency provide anindication of the overall level of appraisal for anygroup of properties represented by a particular salessample. Point estimates of these measures are calcu-lated as shown in table 1. Reliability statistics shouldalso be calculated around each of these measures (seeSection 8.4.3 on Measures of Reliability). Commonmeasures of appraisal level include the mean, salesweighted mean, and median ratios.

8.4.2 Measures of VariabilitySeveral statistical tests are available and should be usedto determine the degree of variability (uniformity) in theproducts of any AVM model. Common measures ofappraisal variability include the coefficient of dispersion(COD) and coefficient of variation (COV).

8.4.2.1 Coefficient of DispersionThe most useful measure of variability is the COD,which measures the average percentage deviation of theratios from the median ratio and is calculated by (1)subtracting the median from each ratio, (2) taking theabsolute value of the calculated differences, (3) sum-ming the absolute differences, (4) dividing by thenumber of ratios to obtain the “average absolute devia-tion,” (5) dividing by the median, and (6) multiplying by100. For the data in table 1:

Average Absolute Deviation =

9.271 ÷ 36 = 0.2575;

COD = (0.2575 ÷ 0.864) * 100 = 29.8.

The COD has the desirable feature that its interpretationdoes not depend on the assumption that the ratios arenormally distributed. Standards for interpreting CODsare contained in Section 14.2 of the Standard on Ratio

Studies (IAAO 1999). Note that the COD represents themean (not the median) percent deviation from themedian. In general, more than half the ratios will fallwithin one COD of the median.

The COD should not be calculated about the meanbecause the mean is more affected by extreme ratiosthan the median, and because of the inherent (upward)bias of the mean of a set of ratios. The COD also shouldnever be calculated about the weighted mean, whichimplicitly weights each ratio based on its sale price.

(IAAO 1999, 24.)

8.4.2.2 Coefficient of VariationThe COV can be another important measure of appraisalvariability. The COV for a sample is calculated by (1)subtracting the mean from each ratio, (2) squaring thecalculated differences, (3) summing the squared differ-ences, (4) dividing by the number of ratios less one toobtain the “variance,” (5) taking the square root to obtainthe “standard deviation,” (6) dividing by the mean, and(7) multiplying by 100. Note that the COV is calculatedonly about the mean—not the median or weighted mean(although other methods permit calculation about theweighted mean). For the data in table 2:

Variance = 3.0808 ÷ 35 = 0.0880;

Standard Deviation = sqrt 0.0880 = 0.2966;

COV = (0.2966 ÷ 0.900) * 100 = 33.0.

The interpretation of the standard deviation and COVrests on the assumption that the ratios are normallydistributed. When this is the case, approximately 68percent of the predicted ratios in the population will liewithin one standard deviation of the mean, and approxi-mately 95 percent will lie within two standard deviationsof the mean. When the ratios do not approximate anormal distribution, these relationships no longer hold(although there always will be at least 75 percent of theratios in any population within two and at least 89percent of the ratios within three standard deviations ofthe mean). Hence, one should determine whether ratiosare approximately normally distributed before using theCOV. When the normality assumption is met, the COVprovides the most precise measure of variability.

Because the deviations between each ratio and the meanratio are squared in determining the COV, ratios thatdiffer greatly from the mean influence the COV morethan they do the COD, in which the deviation of eachobservation from the median is equally weighted.

(IAAO 1999, 25.)

8.4.3 Measures of ReliabilityReliability, in a statistical sense, concerns the degree ofconfidence one can place in a calculated statistic for a


26

Table 1. Example of Ratio Study Statistical Analysis

Data analyzedData analyzedData analyzedData analyzedData analyzed

Rank of ratio of observationRank of ratio of observationRank of ratio of observationRank of ratio of observationRank of ratio of observation Appraised value (AV in $)Appraised value (AV in $)Appraised value (AV in $)Appraised value (AV in $)Appraised value (AV in $) Market value (MV in $)Market value (MV in $)Market value (MV in $)Market value (MV in $)Market value (MV in $) Ratio (AV/MV)Ratio (AV/MV)Ratio (AV/MV)Ratio (AV/MV)Ratio (AV/MV)

1 48,000 138,000 0.3482 28,800 59,250 0.4863 78,400 157,500 0.4984 39,840 74,400 0.5355 68,160 114,900 0.5936 94,400 159,000 0.5947 67,200 111,900 0.6018 56,960 93,000 0.6129 87,200 138,720 0.629

10 38,240 59,700 0.64111 96,320 146,400 0.65812 67,680 99,000 0.68413 32,960 47,400 0.69514 50,560 70,500 0.71715 61,360 78,000 0.78716 47,360 60,000 0.78917 58,080 69,000 0.84218 47,040 55,500 0.84819 136,000 154,500 0.88020 103,200 109,500 0.94221 59,040 60,000 0.98422 168,000 168,000 1.00023 128,000 124,500 1.02824 132,000 127,500 1.03525 160,000 150,000 1.06726 160,000 141,000 1.13527 200,000 171,900 1.16328 184,000 157,500 1.16829 160,000 129,600 1.23530 157,200 126,000 1.24831 99,200 77,700 1.27732 200,000 153,000 1.30733 64,000 48,750 1.31334 192,000 144,000 1.33335 190,400 141,000 1.35036 65,440 48,000 1.363

Note: Due to rounding, totals may not add to match those on following table, which reports results of statistical analysis of above data.

Results of statistical analysisResults of statistical analysisResults of statistical analysisResults of statistical analysisResults of statistical analysis

StatisticStatisticStatisticStatisticStatistic Result calculated on preceding dataResult calculated on preceding dataResult calculated on preceding dataResult calculated on preceding dataResult calculated on preceding data

Number of observations in sample 36Total appraised value $3,627,040Total market value $3,964,620Average appraised value $100,751Average market value $110,128Mean ratio 0.900Median ratio 0.864Geometric mean ratio 0.849Weighted mean ratio 0.915Price-related differential (PRD) 0.98Coefficient of dispersion (COD) 29.8%Standard deviation 0.297Coefficient of variation (COV) 33.0%Probability that population mean ratio is

between 90% and 110% 49.7%95% mean two-tailed confidence interval 0.799–1.00095% median two-tailed confidence interval 0.684–1.06795% weighted mean two-tailed confidence interval 0.806–1.024Shape of distribution of ratios Normal (based on binomial distribution)Date of analysis 9/99/9999Category or class being analyzed Residential

27


sample (for example, how accurately does the samplemedian ratio approximate the true [population] medianappraisal ratio?). There are two related measures ofreliability: confidence intervals and standard errors. Aconfidence interval consists of two numbers that bracketa calculated measure of central tendency for the sample;one can have a specified degree of confidence that thetrue measure of central tendency for the population fallsbetween the two numbers. Standard errors relate to thedistance one must add to and subtract from certainmeasures of central tendency to compute the confi-dence interval.

For the data in table 1, the 95 percent confidence intervalfor the median is 0.684 to 1.067 (calculations notshown)—from the sample data, one can be ninety-fivepercent confident that the median level of appraisal forthe population is in this range. Although most commonlycalculated around the mean, confidence intervals can becalculated about various measures of appraisal level andvariability, or about a resulting property value estimate;standard errors can be properly calculated about themean and weighted mean, or about an estimate of valuefor the population. (See IAAO [1990, 515–546] andGloudemans [1999, 257–339] for information on per-forming these calculations.) The article, “ConfidenceIntervals for the COD: Limitations and Solutions”(Gloudemans 2001), provides criteria for evaluatingwhether CODs can be deemed to have exceeded stan-dards.

Measures of reliability explicitly take into account theerrors inherent in a sampling process. In general, thesemeasures will be tighter (better) when samples arerelatively large and the uniformity of ratios is relativelygood. Although the mathematics of calculating thesemeasures is comparatively straightforward, their cor-rect interpretation is critical and requires someone wellgrounded in the underlying statistical principles.

Users must give careful consideration to reliability mea-sures in evaluating AVM output.

(IAAO 1999, 25.)

8.4.4 Vertical InequitiesThe COD and COV relate to “horizontal,” or random,dispersion among the ratios in a stratum, regardless ofthe value of individual parcels. Another form of inequitymay be systematic differences in the appraisal of low-value and high-value properties, termed “vertical”inequities. When low-value properties are appraised atgreater percentages of market value than high-valueproperties, appraisal regressivity is indicated. Whenlow-value properties are appraised at smaller percent-ages of market value than high-value properties, appraisalprogressivity is the result. Appraisals should be neitherregressive nor progressive.

An index statistic for measuring vertical equity is thePRD (Price-Related Differential), which is calculatedby dividing the mean by the weighted mean:

Mean/Weighted Mean =

Price-Related Differential

This statistic should be close to 1.00. Measures significantlyabove 1.00 tend to indicate appraisal regressivity; measuresbelow 1.00 suggest appraisal progressivity. For the data intable 1, the PRD is 0.983, suggesting slight progressivity.When samples are small or the weighted mean is heavilyinfluenced by several extreme sales prices, however, thePRD may not be a reliable measure of vertical inequities. Ifnot representative, extreme sales prices may be excluded incalculation of the PRD. Similarly, when samples are verylarge, the PRD may be too insensitive to show small pocketsof properties in the population where there is significantvertical inequity.

(IAAO 1999, 26.)

8.4.5 Guidelines for Evaluation of QualityBecause the development and utilization of automatedvaluation models are ongoing, without definitive begin-ning or end dates, sales ratio studies should be performedon a scheduled, periodic basis to establish the currentperformance status of the model. Such ratio studiesshould be conducted utilizing holdout samples accumu-lated according to Section 8.7. Model accuracy shouldbe measured against the Standard on Ratio Studies(IAAO 1999) for the particular property type valued bythe model. The Standard on Ratio Studies (IAAO 1999)suggests that the level of AVM estimate-to-sale price ineach stratum (group of like properties) should be within5 percent of the overall estimate-to-sale ratio for allstrata; and the overall estimate-to-sale level should bewithin 10 percent of the desired level of 100 percent. Forresidential properties, variability, as measured by thecoefficient of dispersion (average percent of error aboutthe median estimate-to-sale price ratio), should be 15percent or less in older, heterogeneous areas and 10percent or less in areas of newer and fairly similarresidences. Variability within strata composed of in-come-producing properties requires a coefficient ofdispersion of 15 percent or less in larger, urban areas,and 20 percent or less in small or rural areas. Within allother types of property strata, the coefficient of disper-sion should be 20 percent or less.

Table 2 is taken from the IAAO Standard on RatioStudies (IAAO 1999) and provides guidelines for evalu-ating the quality of appraisal level and variability basedon statistical measures previously discussed.

8.4.6 Importance of Sample SizeThere is a general relationship, between statistical pre-cision and the number of observations in a sample,


28

drawn from a given population: the larger the sample, thegreater the precision. The required sample size for anygiven degree of precision depends primarily on acceptablesampling error and the variability in the population. Whenthere are insufficient sales to achieve target levels ofprecision, all valid sales should be used unless this resultsin nonrepresentativeness. If an abundance of sales isavailable, it is permissible to randomly include sufficientsales to obtain uniform or reasonably small margins oferror.

Table 3 demonstrates the relationship between sample sizerequirements and variability as measured by the COV withthe values in the table indicating margins of error that mustbe added to and subtracted from the sample mean todetermine the confidence intervals. For example, a sampleconsisting of ten sales with a COV of 20 percent wouldproduce a 95 percent confidence interval with a width of±14.3 percent around the mean. Given the same COV witha sample size of 100 sales, the 95 percent confidenceinterval width would be reduced to ±3.9 percent around themean, thus providing greater precision.

8.5 Property IdentificationAVM developers must accurately identify property in orderto produce an accurate valuation estimate for that property.The common property identification for the commercialAVM industry has become the property address. However,third party data providers use different variations of ad-dresses. Many assessment jurisdictions have not fullystandardized their addresses. Some condominium com-plexes have the same street address for all units.Condominium unit numbers assigned by the assessmentjurisdiction may be the postal number or the lot number ofthe subdivision. These are just some of the variations inaddresses that causes errors or misidentification of theproperties requested by AVM users.

AVM developers attempt to minimize property identifi-cation errors by using address standardization softwarefor all data to be used in the AVM system. All electronic

real estate property systems should move to standard-ized addressing systems such as the Coding AccuracySupport System (CASS) certified by the United StatesPostal Service. While CASS is a way to standardizeaddresses across the U.S., it is primarily intended toensure the accuracy of addresses for mail deliverypurposes. This is a slightly different goal than theidentification of the physical location of the property,especially in rural areas.

One precondition of address standardization is parsing theaddress into separate fields for the number, directional,street name/number, prefixes, and suffixes. Once this isaccomplished, the correction or standardization of theaddress can begin. For example, Florida may be representedby the word Florida or abbreviations such as “Fla.” or “FL.”

Geographic information systems can be used to matchAVM system property addresses to addresses in theU.S. Census Tiger files (or enhanced Tiger files pro-vided by third parties). These GIS files have identified/located addresses by latitude and longitude at the streetaddress segment level for most of the United States.Other countries have similar methods to geocode ad-dresses to locational reference systems.

AVM users also have a responsibility to provide accurateaddresses when requesting an AVM report. They shouldreview the returned AVM report to confirm that thevalue estimate is for the property in question. Valid AVMreports are important for measuring the quality of theAVM system. This is called the hit rate, which is ameasure of the number of usable AVM valuation reportscompared to the total number of valuation reportsrequested. The hit rate will vary by several factors suchas address mismatch; missing data within the propertyrecord that prevents the estimation of value; type ofproperty is outside the scope of the AVM model; and thesize or valuation of the subject property is outside therange of acceptable quality as determined by the qualityassurance review of the model.

Table 2. Ratio Study Performance Standards

Type of propertyType of propertyType of propertyType of propertyType of property Measure of central tendencyMeasure of central tendencyMeasure of central tendencyMeasure of central tendencyMeasure of central tendency CODCODCODCODCOD PRD*PRD*PRD*PRD*PRD*

Single-family residentialNewer, more homogenous areas 0.90–1.10 10.0 or less 0.98–1.03Older, heterogeneous areas 0.90–1.10 15.0 or less 0.98–1.03Rural residential and seasonal 0.90–1.10 20.0 or less 0.98–1.03

Income-producing properties 0.90–1.10Larger, urban jurisdictions 0.90–1.10 15.0 or less 0.98–1.03Smaller, rural jurisdictions 0.90–1.10 20.0 or less 0.98–1.03

Vacant land 0.90–1.10 20.0 or less 0.98–1.03Other real and personal property 0.90–1.10 Varies with local conditions 0.98–1.03

*The standards for the PRD are not absolute when samples are small or when wide variations in prices exist. In suchcases, appropriate tests are more useful (see table 5 of the Standard on Ratio Studies [IAAO 1999, 27]).

29


(Collateral Risk Management Consortium (CRC) 2003, 6.)

8.6 OutliersThe term “outliers” is defined in the Glossary forProperty Appraisal and Assessment (IAAO 1997) asobservations that have unusual values; that is, theydiffer markedly from a measure of central tendency.Some outliers occur naturally; others are due to dataerrors. In valuation models, outliers may include parcelswith unusual characteristics as well as those withextreme estimated values per unit. Large, difficult toexplain differences with respect to previous or controlmodel runs may also identify outliers. Failure to under-stand and address outlier influences may result inunstable models that produce unpredictable changes invalue over time. Documentation accompanying theautomated valuation model must describe the methodol-ogy used to identify outliers and the procedures/trimmingcriteria followed once outliers are identified.

In ratio studies, outlier ratios are very low or high ratios ascompared with other ratios in the sample. When the sampleis small, outlier ratios may distort calculated ratio studystatistics. Some statistical measures, such as the medianratio, are resistant to the influence of outliers. However, theCOD and mean are sensitive to extreme ratios.

Outliers in AVM models can result from any of thefollowing:

1. an erroneous sale price

2. a nonmarket sale

3. unusual market variability

4. a mismatch between the property sold andthe property appraised

5. an error in the appraisal of an individual parcel

6. an error in the appraisals of a subgroup

(IAAO 1999, 19–20.)

One extreme outlier can have controlling influence oversome statistical measures. Particular care must be takento identify outliers if point estimates are used to makeinferences about population level or variability. If, after

proper verification, screening, and editing, an outlierwith a nonrepresentative ratio remains in a study,statistical results will not reflect population level andvariability. The potential distortion is greater whensample size is small. If outliers can be identified, trim-ming procedures are acceptable methods for creating amore representative sample. One outlier identificationmethod is based on the interquartile range; however,because of the skewed distribution of ratios, this proce-dure may locate only extremely high ratios. If one or twohigh outlier ratios are trimmed from a small sample, thestatistical measures of level may be shifted significantlylower. (See Tomberlin [1997] and Hoaglin, Mosteller,and Tukey [1983] on trimming small samples.)

(IAAO 1999, 20.)

8.7 Holdout SamplesHoldout samples represent groups of valid sales selected ina manner that guarantees their group characteristics matchthose of the population of properties covered by theautomated valuation model. Such samples should be accu-mulated at the same time sales are collected for modelcalibration, but used for testing the calibrated model.Inherent in the definition of holdout samples is the premisethat the sales not be used in developing the original model.Sales that occur after model calibration can also be used intesting and validating the model, and this method may bepreferable when few sales are available.

8.8 Value ReconciliationWhen a model is designed to produce more than onevalue estimate for a subject property, model documen-tation must contain a thorough explanation of theprocedures followed to reconcile those candidate esti-mates into a final estimate of value. Those proceduresmust include analysis of the relative strengths andweaknesses of the candidate estimates, and specifica-tion of how that analysis results in a final value estimate.

In those instances in which all candidate estimates arepresented to the user for their reconciliation, the systemmust report the quantity and quality of data supportingeach of the candidate estimates. If the product of anautomated valuation model is a set of value estimatesderived from more than one of each of the threeapproaches, that product must also include sufficientinformation to allow the user to weigh the validity ofthose estimates, based on the quality and quantity of dataavailable to support them.

When the model is designed to produce estimates ofvalue for individual properties, those estimates must beaccumulated and compared to their actual selling pricesusing ratio studies conducted at regular intervals. Inaddition, confidence intervals can be calculated aroundvalue estimates developed for individual parcels. Nar-row intervals indicate greater likelihood that the estimate

Table 3. Confidence Intervals and Sample Size:95 Percent Confidence Interval

Sample sizeSample sizeSample sizeSample sizeSample size COV = 10.0COV = 10.0COV = 10.0COV = 10.0COV = 10.0 COV = 20.0COV = 20.0COV = 20.0COV = 20.0COV = 20.0 COV = 30.0COV = 30.0COV = 30.0COV = 30.0COV = 30.0

5 ±12.4 ±24.8 ±37.210 ±7.2 ±14.3 ±21.550 ±2.8 ±5.5 ±8.3

100 ±2.0 ±3.9 ±5.9300 ±1.1 ±2.3 ±3.4


30

reflects market value. Additionally, z scores can becalculated and show the number of standard deviationsby which an AVM estimated value misses actual salesprice. Properties with value misses outside of a ±3standard deviation range should be reviewed for system-atic model error.

8.9 Appraiser Assisted AVMsWhen an appraiser reviews or changes an AVM reportprepared by a separate AVM provider, the results arecalled appraiser-assisted AVMs (AAVMs). The ap-praiser can provide an additional opinion of the estimatedvalue and usually will sign the report and confirm thevalue. All AVM reports can have their estimates of valueoverridden by an appraiser’s opinion of value. In mostcases, appraisers are limited in their ability to change anAVM report. AVM reports based on the traditionalformats of the cost, sales comparison and incomeapproaches are the easiest for appraisers to change.

8.10 Frequency of UpdatesAVM estimates of value are based on formulas derivedfrom market analysis of a specific geoeconomic areaduring a specified time frame. Because AVM valueestimates represent trends in time as applied to a specificproperty with known characteristics (physical and/oreconomic), AVM providers must update their formulas,estimates of value, characteristics, and economic data-bases regularly. Movements in the market and theavailability of market information should dictate thefrequency of this process.

9. AVM REPORTSThere are three general types of reports that are consid-ered part of the AVM reporting process. They are thedetailed documentation report, the restricted use report,and the appraiser-assisted report. In all cases, thereports should be in compliance with the respectiveportions of USPAP.

9.1 Types of ReportsThere are several report formats associated with thedevelopment of an AVM and the reporting of an indi-vidual property’s estimate of value. Documentationreports, restricted use reports, and CAMA/AAVM re-ports each provide different reporting levels of appraisalanalysis within the report.

9.1.1 Documentation ReportThere are several report formats associated with thedevelopment of an AVM and the reporting of anindividual property’s estimate of value. The develop-ment of an AVM formula involves the analysis of thehistorical market place (real estate) information inorder to create value estimates at a particular point intime. This market analysis should comply with USPAP

Standard 6: Mass Appraisal, Development and Re-porting (Appraisal Foundation 2003, 46–56). Thereshould be a detailed report to document and supportthe market analysis process and the final valuationformula. This includes the sections of USPAP Stan-dard Rule 6–7 (report format) and 6–8 (certification)(Appraisal Foundation 2003, 53–55).

9.1.2 Restricted Use ReportWhen requesting an AVM, the client is normally notinterested in complete narrative reports as described inStandard 2: Real Property Appraisal Reporting, or Stan-dard 6: Mass Appraisal, Development and Reporting(Appraisal Foundation 2003, 21–31, 46–56). AVMclients want quick standardized indicators of value, thatmay be retrieved from the AVM systems by supportpersonnel without professional real estate training orknowledge. This includes the general public, which maybe interested in an indication of value for properties thatthey already know, such as property owners whorequest an AVM to check the current market valuebefore making various economic decisions. This re-quires a restricted use report that is limited to theimmediate intended user (client) of the AVM report.These restricted use reports are typically limited togenerally acceptable property identification such asstreet address, indication of value, some basic prop-erty descriptive characteristics, known additionalindicators of value (such as last sale price/date andproperty tax assessment), and report date. Theremay be additional qualification and limiting condi-tions information as described in USPAP Standard6–7 (mass appraisal report) (Appraisal Foundation2003, 53–55) that is not of general interest to theintended user. These restricted reports are generallyone to a few pages in length. These are the reportsreferred to in USPAP AO–18, Use of an AutomatedValuation Model (AVM) (Appraisal Foundation 2003,180-187), which states that the output of an AVM isnot, by itself, an appraisal. These restricted usereports are simply the application of an AVM modelformula to an individual property and do not containthe supporting documentation of the appraisal pro-cess performed to create the formula, which shouldbe in the documentation report.

9.1.3 CAMA or AAVM ReportA third type of report is the combination of AVMformulas with appraisers’ review and verification ofvaluations. These are sometimes called CAMA in gov-ernment tax assessment and appraiser-assisted AVMreports, in the commercial AVM field. This type ofreport combines the most desirable parts of the AVM(unbiased market analysis and consistently applied modelformulas) with the most desirable parts of the fieldappraiser (property inspection, local knowledge and

31


experience). The AVM provider sends the AVM reportto the appraiser in electronic format. The appraiserperforms a desktop review or one of various levels ofinspection, as desired by the client, and corrects/con-firms the AVM report and value estimate before deliveryof the appraiser’s final opinion of value to the client.

While all AVM reports can have their estimates of valueoverridden by an appraiser’s opinion of value, in mostcases, appraisers are limited in their ability to change thecomparable selections, calculations, and variable adjust-ments within an AVM report. AVM reports based on thetraditional formats of the cost, sales comparison, andincome approaches, are the easiest for appraisers tochange or adjust at the individual variable level.

9.2 Uses of AVMAVM reports may have many uses. This standard willonly list some of the typical uses.

9.2.1 Real Estate Lenders• Reduce time to approve real estate loan

applications

• Provide unbiased estimate of value for loanunderwriting

• Provide real estate value/scores to complimentborrower’s credit scoring

• Standard estimates for annual review ofindividual appraiser’s performance

• Quality assurance for selling pooled loans

• Review of loan portfolios

• Support for lending decisions and geographicdistribution required by the CommunityReinvestment Act

• Statistical support for litigation

• Updates current valuation of portfolio properties

• Support in purchase of loan portfolios orlending institutions

• Portfolio valuation reviews by secondarymortgage markets and bond rating firms

• Systematic review of mortgage loan transactionto assist in the discovery of potential fraud

9.2.2 Real Estate Professional• Support in setting listing price

• Support in negotiation between sellers andbuyers

• Central database for appraisers

• Support for appraiser’s opinions of value

• Support for appraiser’s review and desktop

appraisal assignments

• Support for appraisal consulting assignmentsthat involve large numbers of properties

• Statistical support for litigation

9.2.3 Government• Planning and land use decisions

• Development of value estimates for review byassessment staff appraisers

• Standardized estimates of value to annuallyreview field appraisers’ performances

• Valuation substitutes for appraisals in ratiostudy reports

• Screening of sale prices for valid market salestransactions

• Audits of lenders by state and federalregulators

• Assist states with standardized values to reviewproperty assessments in school fundingformulas

• Fraud identification and prevention byenforcement, taxation, customs, and oversightagencies (such as GSE, HUD, IRS, CanadaMortgage and Housing Corporation, StatisticsCanada, and state and national bank regulators)

• Fraud prosecution by comparing transactionsto standardized values

• Assist in valuation for right-of-way andproperty condemnation cases

9.2.4 General Public• Support for various business development and

economic decisions

• Assistance in determining best listing price

• Assistance in determining best offering price

• Review of local government tax assessments

• Estate estimates of real estate value byattorneys and estate administrators

AVM reports may be sufficient as stand-alone products, orthey may lead to a request for a more detailed appraisalreport based on the needs and usage of the intended user.This listing is only a portion of the potential uses of AVMs.When clients request AVMs for a limited and specific use,the AVM report will provide quality information to theintended user quickly and inexpensively.

10. GLOSSARYAlgorithm—Computer-oriented, precisely defined set of


32

steps that, if followed exactly, will produce a prespecifiedresult (for example, the solution to a problem).

Additive Model—A model in which the dependent variableis estimated by multiplying each independent variable by itscoefficient and adding each product to the constant.

Appraisal Emulation Model—The appraisal emulationmodel (see Section 3.2.2.1 Comparable Sales Method)follows the steps that an appraiser might follow informing a value estimate (although not with the sameinsight or flexibility that a qualified appraiser brings to theassignment). The model selects “comparable sales”using some standard criteria. It then rates those compa-rable sales by suitability, based on the physical and salescharacteristics of each comparable sale, by adjusting thevarying elements (much as is done on an appraisalform); the model then calculates an estimate of value.

Automated Valuation Model—An automated valuationmodel (AVM) is a mathematically based computersoftware program that produces an estimate of marketvalue based on market analysis of location, marketconditions, and real estate characteristics from informa-tion that was previously and separately collected. Thedistinguishing feature of an AVM is that it is a marketappraisal produced through mathematical modeling.Credibility of an AVM is dependent on the data used andthe skills of the modeler producing the AVM.

Binary (Dummy) Variable—(1) Binary variables arequalitative data items that have only two possibilities—yes or no (for example, corner location). (2) A variablefor which only two values are possible, such as resultsfrom a yes-or-no question; for example, does thisbuilding have any fireplaces? Used in some models toseparate the influence of categorical variables. Alsocalled a dichotomous variable or dummy variable.

Blended Model—A blended model (see Section 8.8: ValueReconciliation) is one where more than one modelingtechnique is used in deriving the estimate of value. Typi-cally, the technique involves running a hedonic model anda repeat sales index. The results are then compared andevaluated. Based on each result, the blended model reportsa final estimate of value. In addition to the hedonic modeland repeat sales index, many blended models also includethe results of a tax-assessed value model.

Calibration—The process of estimating the coeffi-cients in a mass appraisal model.

Coefficient—(1) In a mathematical expression, a num-ber or letter preceding and multiplying another quantity.For example, in the expression “5X”, 5 is the coefficientof X, and in the expression “aY”, a is the coefficient ofY. (2) A dimensional statistic, useful as a measure ofchange or relationship.

Cluster Analysis—A statistical technique for groupingcases (for example, properties) based on specified vari-ables such as size, age, and construction quality. Theobjective of cluster analysis is to generate groupings that areinternally homogeneous and highly different from oneanother. Various cluster algorithms can be employed.

Cost Approach—(1) One of the three approaches to value,the cost approach is based on the principle of substitution–that a rational, informed purchaser would pay no more fora property than the cost of building an acceptable substitutewith like utility. The cost approach seeks to determine thereplacement cost new of an improvement minus deprecia-tion plus land value. (2) The method of estimating the valueof property by: (a) estimating the cost of construction basedon replacement or reproduction cost new, or trendedhistorical cost (often adjusted by a local multiplier); (b)subtracting depreciation; and (c) adding the estimated landvalue. The land value is most frequently determined by thesales comparison approach.

Data Management—The human (and sometimes com-puter) procedures employed to ensure that no informationis lost through negligent handling of records from a file,all information is properly supplemented and up-to-date,and all information is easily accessible.

Direct Market Method/Analysis—One of two formats ofthe sales comparison approach to value (the other being theComparable Sales Method). In the direct market method,the market analyst specifies and calibrates a single modelused to estimate market value directly using multiple regres-sion analysis or another statistical algorithm.

Economic Area—A geographic area, typically encom-passing a group of neighborhoods, defined on the basis thatthe properties within its boundaries are more or less equallysubject to a set of one or more economic forces that largelydetermine the value of the properties in question.

Euclidean Distance Metric—A measure of distancebetween two points “as the crow flies.” In propertyvaluation, it is used to find the nearest neighbor or similarproperty based on an index of dissimilarity betweenproperty location or attributes. When using multivariateselection, the squared difference is divided by thestandard deviation of the variable so as to normalize thedifferences. (Also see Minkowski Metric.)

Hedonic Model—Hedonic pricing attempts to take ob-servations of the overall goods or services and obtainimplicit prices for the goods and services. Prices aremeasured in terms of quantity and quality. When valuingreal property, the spatial attributes and property-specificattributes are valued in a single model. Calibration of theattribute components is performed statistically by re-gressing the overall price onto the characteristics.

Heteroscedasticity—Nonconstant variance; specifically,in regression analysis, a tendency for the absolute errorsto increase (fan out) as the dependent variable increases.

Holdout Sample—Part of a set of data set aside fortesting the results of analysis.

Homogeneous—Possessing the quality of being alike innature and therefore comparable with respect to theparts or elements; said of data if two or more sets of dataseem drawn from the same population; also said of dataif the data are of the same type (that is, if counts, ranks,and measures are not all mixed together).

Hybrid Model—Model that incorporates both additive

33


and multiplicative components. (See also Additive Model,Hedonic Model, and Multiplicative Model.)Income Approach—One of the three approaches tovalue, based on the concept that current value is thepresent worth of future benefits to be derived throughincome production by an asset over the remainder of itseconomic life. The income approach uses capitalizationto convert the anticipated benefits of the ownership ofproperty into an estimate of present value.

Geographic Information System (GIS)—(1) A data-base management system used to store, retrieve,manipulate, analyze, and display spatial information. (2)One type of computerized mapping system capable ofintegrating spatial data (land information) and attributedata among different layers on a base map.

Goodness-of-Fit—A statistical estimate of the amount,and hence the importance, of errors or residuals for allthe predicted and actual values of a variable. In regres-sion analysis, for example, goodness-of-fit indicateshow much of the variation between independent vari-ables (property characteristics) and the dependent variable(sales prices) is explained by the independent variableschosen for the AVM.

Location Value Response Surface Analysis—A massappraisal technique that involves creating value influencecenters, computing variables to represent distances (ortransformations thereof) from such points and using thevariables in a multiple regression or other model to capturelocation influences. Implementation of the technique isenhanced by the use of a geographic information system.Some geographic information systems permit the valueinfluence centers to be displayed and measured as a three-dimensional grid surface, the results of which can belikewise used in calibration techniques to arrive at thecontribution of location based on the model specification.

Location Variable—A variable that seeks to measurethe contribution of locational factors to the total propertyvalue, such as the distance to the nearest commercialdistrict or the traffic count on an adjoining street.

Market—(1) The topical area of common interest inwhich buyers and sellers interact. (2) The collectivebody of buyers and sellers for a particular product.

Market Analysis—A study of real estate market condi-tions for a specific type of property.

Market Analyst—An appraiser who studies real estatemarket conditions and develops mathematical formulasthat represent those market conditions.

Market Area—(See Economic Area.)

Market Value—Market value is the major focus of mostreal property appraisal assignments. Both economic andlegal definitions of market value have been developedand refined. A current economic definition agreed uponby agencies that regulate federal financial institutions inthe United States is:

The most probable price (in terms of money) whicha property should bring in a competitive and open

market under all conditions requisite to a fair sale, thebuyer and seller each acting prudently and knowl-edgeably, and assuming the price is not affected byundue stimulus. Implicit in this definition is the con-summation of a sale as of a specified date and thepassing title from seller to buyer under conditionswhereby:

• The buyer and seller are typically motivated;

• Both parties are well informed or well advised,and acting in what they consider to be theirbest interests;

• A reasonable time is allowed for exposure inthe open market;

• Payment is made in terms of cash in UnitedStates Dollars or in terms of financialarrangements comparable thereto.

The price represents the normal consideration for theproperty sold unaffected by special or creative financingor sales concessions granted by anyone associated withthe sale.

Mean—A measure of central tendency. The result ofadding all the values of a variable and dividing by the numberof values. For example, the mean of three, five, and ten, istheir sum (eighteen) divided by three, which is six.

Median—A measure of central tendency. The value ofthe middle item of an uneven number of items arrangedor arrayed according to size; the arithmetic average ofthe two central items in an even number of itemssimilarly arranged.

Minkowski Metric—Any of a family of possible ways ofmeasuring distance. Euclidean distance, a member of thisfamily, computes straight-line distances (as the crow flies)by squaring differences in like coordinates, summing them,and taking the square root of the sum. In mass appraisalmodel building, Minkowski metric usually refers to the sumof absolute differences (not squared) in each dimension,and resembles a “taxicab” or city block pattern. Otheralternatives are possible, including the distance as calcu-lated only for the dimension of greatest difference, but thecity block distance is most common.

Model—(1) A representation of how something works.(2) For purposes of appraisal, a representation (in wordsor an equation) that explains the relationship betweenvalue or estimated sale price and variables representingfactors of supply and demand.

Model Specification—The formal development of amodel in a statement or equation, based on data analysisand appraisal theory.

Model Calibration—The development of the adjust-ments or coefficients from market analysis of thevariables to be used in an automated valuation model.

Multicollinearity—Correlation among two or more vari-ables. In regression analysis, high multicollinearity amongthe independent variables complicates modeling and willcompromise the reliability of the resulting coefficients.


34

If the multicollinearity is perfect, the multiple regressionalgorithms simply will not work and either an errormessage may result or the software may purge one ormore of the problem variables.

Multiplicative Model—A mathematical model in whichthe coefficients of independent variables serve as pow-ers (exponents) to which the independent variables areraised, or in which independent variables themselvesserve as exponents; the results are then multiplied toestimate the value of the dependent variable.

Multiple Regression Analysis (MRA)—A particularstatistical technique, similar to correlation, used toanalyze data in order to predict the value of one variable(the dependent variable), such as market value, from theknown values of other variables (called “independentvariables”), such as lot size, number of rooms, and soon. If only one independent variable is used, the proce-dure is called simple regression analysis and differs fromcorrelation analysis only in that correlation measures thestrength of the relationship, whereas regression predictsthe value of one variable from the value of the other.When two or more variables are used, the procedure iscalled multiple regression analysis.

Neighborhood—(1) The environment of a subject prop-erty that has a direct and immediate effect on value. (2)A geographic area (in which there are typically fewerthan several thousand properties) defined for someuseful purpose, such as to ensure for later multipleregression modeling that the properties are homoge-neous and share important locational characteristics.

Neighborhood Analysis—A study of the relevant forcesthat influence property values within the boundaries ofa homogeneous area.

Neural Network—An artificial neural network (ANN) isa collection of mathematical models that emulate someof the observed properties of biological nervous systemsand draw on the analogies of adaptive biological learning.An artificial neural network has several key elements:input, processing (calibration), and output. Other namesassociated with neural networks include: connect-ionism,parallel distributed processing, neuro-computing, natu-ral intelligent systems, and machine learning algorithms.

Outlier—An observation that has unusual values, that is, itdiffers markedly from a measure of central tendency. Someoutliers occur naturally; others are due to data errors.

Ratio Study—A study of the relationship betweenappraised or assessed values and market values. Indica-tors of market values may either be sales (sales ratiostudy) or independent “expert” appraisals (appraisalratio study). Of common interest in ratio studies are thelevel and uniformity of the appraisals and assessments.

Repeat Sales Analysis Model—Repeat sales analysis(see Section 4.4: Time Series Analysis) aggregateschanges in value and statistical means for propertiessold more than once during a specified period of timein a given geographic area. For example, in a zip orpostal code area, estimate market-level housing price

changes. If an individual property has not beensubstantially changed since its last sale, this analysismatches each pair of sales transactions (thus thename “repeat sales”). The amount of appreciation(or depreciation) is calculated from the time of thefirst sale to the second and so on, providing anestimate of the overall appreciation of that localhousing market during that time period.

The larger the number of available sales pairs, the morestatistically reliable the estimate of overall housing pricetrends will be. Because this analysis is based on identi-fying properties where more than one sale has occurred,the challenge is to identify enough observations toprovide a meaningful index of housing values, whilekeeping to as small a geographic area as possible.

A repeat sales index may also overestimate marketappreciation if the data contains pairs of sales in whichthe second sales price reflects substantial improvements(or other alterations) made to the property after the firstsale. On the other hand, repeat sales indices can and doprovide very useful valuation estimates in jurisdictionswhere the data is insufficient to support hedonic models.In addition, they may prove to be more accurate intracking housing values for the houses that a hedonicmodel may struggle with (especially those subject toextreme positive or negative influences) when a priorsale is known on the property.

Sales Comparison—One of the three approaches tovalue, the sales comparison approach estimates aproperty’s value (or some other characteristic, such asits depreciation) by reference to comparable sales.

Stepwise Regression—A kind of multiple regressionanalysis in which the independent variables enter themodel, and leave it, if appropriate, one by one accordingto their ability to improve the equation’s power to predictthe value of the dependent variable.

Software—Anything that is stored electronically on acomputer is software. The storage device is hardware.There are two general categories of software: (a) oper-ating systems and the utilities that allow the computer tofunction, and (b) applications which are programs thatallow users to work with the computer (e.g., wordprocessing, spreadsheets, databases, AVMs).

Stratification—The division of a sample of observa-tions into two or more subsets according to somecriterion or set of criteria. Such a division may be madeto analyze disparate property types, locations, or char-acteristics, for example.

Tax Assessed Value Model—Tax assessed value modelsderive an estimate of value by examining market valuesattributed to properties by the local taxing authorities (seeSection 4.5 Tax Assessed Value Model). As a matter of locallaw and custom, the values reported by the taxing authoritiesoften (but not always) vary from the current market value insome reasonably predictable manner. For example, somejurisdictions require the taxing authority to report the value at 25percent of estimated market value. In others, values are re-

35


assessed only on an infrequent basis. Some jurisdictions reportmultiple values—assessed, appraised and market values. Byexamining local laws and customs with respect to how thatvalue is derived, it is often possible to provide a generaladjustment to values reported by taxing authorities to betterapproximate current market value.

Time Series Analysis—A family of techniques that canbe used to measure the cyclical movements, randomvariations, seasonal variations, and secular trends ob-served over a period of time.

Weighted Mean—An average in which each value isadjusted by a factor reflecting its relative importance inthe whole, before the values are summed and divided bytheir number.

Variable—An item of observation that can assumevarious values, such as square feet, sales prices, or salesratios. Variables are commonly described using mea-sures of central tendency and dispersion.

ReferencesAmerican Planning Association. 2003. “Land BasedClassification Standards.” Available fromwww.planning.org/LBCS/GeneralInfo.

Appraisal Foundation. 2003. Uniform standards of pro-fessional appraisal practice (USPAP). Washington,D.C.: Appraisal Foundation.

Appraisal Institute. 2002. The dictionary of real estateappraisal. 4th ed. Chicago: Appraisal Institute.

Carbone, R. 1976. The design of an automated massappraisal system using feedback. PhD diss., Carnegie-Mellon University.

Collateral Risk Management Consortium (CRC). 2003.The CRC guide to automated valuation model (AVM)performance testing. Paper presented at Fidelity Na-tional Information Solutions (FNIS) Valuation Innovationand Leadership Summit, CRC, May 28, in LagunaBeach, CA.

D’Agostino, R.B., and Stephens, M.A. 1986. Good-ness-of-fit techniques. New York: Marcel Dekker.

Gloudemans, R. J. 1999. Mass appraisal of real prop-erty. Chicago: IAAO.

Gloudemans, R.J. 2001. Confidence intervals for thecoefficient of dispersion: Limitations and solutions.Assessment Journal 8 (6):23-27.

Guerin, B.G. 2000. MRA model development usingvacant land and improved property in a single valuationmodel. Assessment Journal 7 (4):27-34.

Hoaglin, D.C., Mosteller, F., and Tukey, J.W. 1983.Understanding robust and exploratory data analysis.New York: John Wiley & Sons.

IAAO. 1990. Property appraisal and assessment admin-

istration. Chicago: IAAO.

IAAO. 1997. Glossary for property appraisal and as-sessment. Chicago: IAAO.

IAAO. 1999. Standard on ratio studies. Chicago: IAAO.

IAAO. 2002. Automated valuation modeling: A primer.Typescript.

Mendenhall, W., and Sincich, T. 1996. A second coursein statistics: Regression analysis. 5th ed. Upper SaddleRiver, NJ: Prentice Hall.

Tomberlin, N. 1997. “Trimming outlier ratios in smallsamples.” Paper presented at IAAO International Con-ference on Assessment Administration, September14-17, at Toronto, ON, Canada.

Waller, B.D. 1999. The impact of AVMs on the appraisalindustry. The Appraisal Journal 67 (3):287-292.

Ward, R.D., and Steiner, L.C. 1988. A comparison offeedback and multivariate nonlinear regression analysisin computer-assisted mass appraisal. Property Tax Jour-nal 7 (1):43-67.

Wollery, A., and Shea, S. 1985. Introduction to com-puter assisted valuation. Boston, MA: Oelgeschlager,Gunn & Hain, Publishers, Inc.

Additional Suggested ReadingsGloudemans, R.J. 2002. Comparison of three residentialregression models: Additive, multiplicative, and nonlin-ear. Assessment Journal 9 (4):25-36.

O’Connor, P.M. 2002. Comparison of three residentialregression models: Additive, multiplicative, and nonlin-ear. Assessment Journal 9 (4):37-44.

Gloudemans, R.J, and O’Connor, P.M. 2002. “Com-parison of three residential regression models: Additive,multiplicative, and nonlinear.” Papers presented at 6thAnnual Integrating GIS/CAMA Conference. Urban andRegional Information Systems Association and IAAO,April 7-10, at Reno, NV. Available on CD-ROM.

Maloney, J., Ripperger, R., and O’Connor, P.M. 2001.“The first application of modern location adjustments tocost approach and its impact.” Linking Our Horizons,67th International Conference on Assessment Adminis-tration. IAAO, September 9-12, at Miami Beach, FL.


36

Assessment Standards of theInternational Association of Assessing Officers

Guide to Assessment Administration Standards ...................................................................February 1990

Standard on Administration of Monitoring and Compliance Responsibilities ................................. July 2003

Standard on Assessment Appeal ................................................................................................ July 2001

Standard on Automated Valuation Models ....................................................................... September 2003

Standard on Digital Cadastral Maps and Parcel Identifiers .......................................................... July 2003

Standard on Contracting for Assessment Services ...............................................................February 2002

Standard on Facilities, Computers, Equipment, and Supplies ............................................ September 2003

Standard on Mass Appraisal of Real Property ....................................................................February 2002

Standard on Professional Development ...............................................................................October 2000

Standard on Property Tax Policy ........................................................................................... August 1997

Standard on Public Relations ...................................................................................................... July 2001

Standard on Ratio Studies .......................................................................................................... July 1999

Standard on the Valuation of Property Affected by Environmental Contamination ......................... July 2001

Standard on Valuation of Personal Property ........................................................................February 1996

To order any standards listed above or to

check current availability and pricing, go to:

http://www.iaao.org/publication/standards.html

Standard on Automated Valuation Models (AVMs)Standard on Automated Valuation Models (AVMs) Approved...

Documents

Transcript of Standard on Automated Valuation Models (AVMs)Standard on Automated Valuation Models (AVMs) Approved...