A New Dynamic Factor Model for Nowcasting Belgian GDP growth · Introduction 2. The analysis of...

RESEARCH & DEVELOPMENT STATISTICS (NBB)

Free and Open Source Software, licensed under the EUPL

(http://ec.europa.eu/idabc/eupl). The last updated version

can be downloaded here https://github.com/jdemetra/jdemetra-app/releases

David de Antonio LiedoJean Palate

Peter Reusens

Time Series Workshop

(Paris 26-27 September 2019)

A New Dynamic Factor Model for

Nowcasting Belgian GDP growth

http://ec.europa.eu/idabc/eupl

https://github.com/jdemetra/jdemetra-app/releases

http://www.ntts2015.eu/

2

NOWCASTING

► Contraction for now and forecasting and has been used for a long-time

in meteorology: 0-6 hours ahead weather forecasts

What is the weather today? Take a look outside the window.

Macroeconomists do not have this luxury. The first official estimate of GDP this

quarter will not be published until the end of January 2017. In fact, we are not even

sure about GDP for Q3

3

NOWCASTING






sure about GDP for Q3 (data is subject to revisions)

-2,50

-2,00

-1,50

-1,00

-0,50

0,00

0,50

1,00

1,50

2,00

20

01Q

1

20

01Q

3

20

02Q

1

20

02Q

3

20

03Q

1

20

03Q

3

20

04Q

1

20

04Q

3

20

05Q

1

20

05Q

3

20

06Q

1

20

06Q

3

20

07Q

1

20

07Q

3

20

08Q

1

20

08Q

3

20

09Q

1

20

09Q

3

20

10Q

1

20

10Q

3

20

11Q

1

20

11Q

3

20

12Q

1

20

12Q

3

20

13Q

1

20

13Q

3

20

14Q

1

20

14Q

3

20

15Q

1

20

15Q

3

20

16Q

1

20

16Q

3

20

17Q

1

20

17Q

3

20

18Q

1

20

18Q

3

20

19Q

1

BE *FLASH BE LAST

4

NOWCASTING






sure about GDP for Q3

► While we wait for these data, we float in a sea of information on all

aspects of the economy: production, sales, economic sentiment, etc…

Nowcasting is about reading the news to figure out if it is rainy or sunny

out there in the economy. This is daily work of economists on trading

desks, at central banks, statistical offices, and in the media

According to the results of a survey conducted by the ECB professional

forecasters admit 40% of their short-term forecast is judgment and 60% comes

from models

5

Outline of the presentation

1. Introduction

2. The analysis of news

• Mapping of news into forecasting updates for euro area growth

• Empirical results of the benchmark case

3. Forecasting Evaluation

• Real-Time Results since 2014

• Reproducing the results with JDemetra+Nowcasting

• Increasing the proportion of Belgian indicators

4. Application: Belgian and euro area GDP growth in 2019Q3

5. Concluding remarks

6

INTRODUCTION

7

Goal of the methodology

► Quantify the real-time impact of survey data on forecasts for GDP growth

Retail Sales

(DE)

Industrial

Production

(DE)

Industrial

Production

(EA)

Sentix GDP (BE)GDP (DE,

EA)

Surveys

Block (IFO,

EC, Markit,

NBB, ZEW)

April

May

June

July

August

September

October

November

December

?

?

…

target variabe, but only

published with a certain delay

8


1. Data revisions

Retail Sales

(DE)

Industrial

Production

(DE)

Industrial

Production

(EA)


EA)

Surveys

Block (IFO,

EC, Markit,

NBB, ZEW)

April

May

June

July

August first estimate first estimate first estimate

September

October

November

December

flash flash

?

?

…


9


1. Data revisions

2. Release schedule (ragged-edge)

Retail Sales

(DE)

Industrial

Production

(DE)

Industrial

Production

(EA)


EA)

Surveys

Block (IFO,

EC, Markit,

NBB, ZEW)

April

May

June

July

August first estimate first estimate first estimate

September

October

November

December

flash flash

?

?

…


10


1. Data revisions

2. Release schedule (ragged-edge), which changes depending on when you

update your series

…Retail Sales

(DE)

Industrial

Production

(DE)

Industrial

Production

(EA)


EA)

Surveys

Block (IFO,

EC, Markit,

NBB, ZEW)

April

May

June

July

August

September new data new data new data

October

November new data

December

flash

yh?


11

Goal: mapping of news into forecasting updates


► Data releases incorporate news: what is the weight of each piece of

news in the forecasts of GDP?

Retail Sales

(DE)

Industrial

Production

(DE)

Industrial

Production

(EA)


EA)

Surveys

Block (IFO,

EC, Markit,

NBB, ZEW)

April

May

June

July

August

September new data new data new data

October

November new data

December

flash

yh?

…

ω4yh

ω5yh

12

Literature

► Lots of research since Giannone, Reichlin and Small (GRS, 2008):

Kalman filter methods to extract the signal from a potentially large

information set and update it in real time.

► Still, many papers consider oversimplistic release schedules or pretend

data revisions do not exist.

► To the best of our knowledge, Basselier et al (2016) are first ones to build

time series composed of press releases for all series (different from the

use of vintages; different from the approach by Kishor & Koening, 2009)

► Here, we extend the model Basselier et al (2016) by including a total

of 80 variables, and we evaluate its forecasting performance

IPI 1 Jan 1 Feb 1 Mar 1 Apr 1 May

2000Q1 100 101 99 99 99

2000Q2 103 102 102 102

2000Q3 106 105 106

2000Q4 108 106

2001Q1 110

13

Literature

► Same methodology as Bańbura and Modugno (2010) : ML estimation of

a DFM à la GRS (2008) to formalize the process of nowcasting:

✓ decompose forecast revisions in terms of news (sample dependent)

✓ assess the expected contribution of each piece of news at forecasting

euro area GDP

(depends on both the properties of the model and the release schedule)

► The model: dynamic factor model with f(t) following a VAR(4)

► We use the nowcasting plugin of JDemetra+

14

ANALYSIS OF NEWS

Summary of the results by

Basselier, de Antonio Liedo and Langenus (2016)

15


1. Introduction










Simple representation of the real-time dataflow

-3

-2

-1

0

1

2

3

4

5

6G

fk S

urv

ey

Ha

rd d

ata

IT

Su

rve

y b

lock

Fla

sh G

DP

BE

Se

ntix S

urv

ey

Ha

rd b

lock

Fla

sh G

DP

fo

reca

st E

A

Fla

sh G

DP

EA

/DE

Gfk

Su

rve

y

Ha

rd d

ata

IT

Su

rve

y b

lock

Se

ntix S

urv

ey

Ha

rd b

lock

Gfk

Su

rve

y

Ha

rd d

ata

IT

Su

rve

y b

lock

Se

ntix S

urv

ey

Ha

rd b

lock

Gfk

Su

rve

y

Ha

rd d

ata

IT

Su

rve

y b

lock

Fla

sh G

DP

BE

Se

ntix s

urv

ey

Ha

rd b

lock

Fla

sh G

DP

fo

reca

st E

A

Fla

sh G

DP

EA

/DE

Reference period Publication date

July

Aug.

Sep.

Oct.

Nov.

Dec.

June

May

April

news2.0

14/08/15news3

01/09/15

news4

16/09/15

news5

01/10/15

news6

16/10/15

news7

01/11/15

news8.0

12/11/15

new

s8.1

13/1

1/1

5

new

s8.2

15/1

1/1

5

new

s2.1

15/0

8/1

5

news1

01/08/15

17

Mapping of news into forecasting updates

► News = actual (published) figure for variable 𝑦 minus the expected value,

conditional on the previous information set

► This definition implies that news cannot be read if we do not have a prior

expectation

► The vector of news (𝐼𝑣+1) can be large, specially if a given release

incorporates historical data revisions

ถℱ𝑣+1𝑟𝑒𝑓𝑟𝑒𝑠ℎ𝑒𝑑

− ดℱ𝑣𝑎𝑟𝑐ℎ𝑖𝑒𝑣𝑒𝑑/𝑜𝑙𝑑

≡ 𝐼𝑣+1 =

y(𝑖,𝑡)1 − E y(𝑖,𝑡)1 ℱ𝑣…

y(𝑖,𝑡)J − E y(𝑖,𝑡)𝐽 ℱ𝑣

18

Quantifying the role of the news

► When updating the nowcasts, the weight of the news depends on

● the quality of the indicator (i.e. the correlation of the factor with the innovations)

● its timeliness (variables that come first will typically have a bigger weight)

[w1𝑘,𝑡 , … , w5

𝑘,𝑡 ] = E[yk,t Iv+1′ ] 𝐸 𝐼𝑣+1 𝐼𝑣+1

′ −1

E[y𝑘,𝑡 ℱ𝑟𝑒𝑓𝑟𝑒𝑠ℎ𝑒𝑑 − E[y𝑘,𝑡 ℱ𝑜𝑙𝑑 =

𝑗=1

𝐽

w𝑗𝑘,𝑡 y(𝑖,𝑡)𝑗 − E y(𝑖,𝑡)𝑗 ℱ𝑜𝑙𝑑

timeliness

quality

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)

Sentix Investor confidence EA

GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE

Industrial production DE

Industrial production EA

ZEW Survey Expectations DE

ZEW Survey Current Situation DE

IFO - Expectations DE

IFO - Business Climate DE

GfK Consumer Confidence DE

Consumer Confidence BE

Business Confidence BE

ZEW Survey Expectations EA

Markit Manufacturing PMI EA

Economic confidence EA

Consumer Confidence EA

Business Climate Indicator EA

EA GDP Q3 (subsequent vintages in %)

Revision

The process of updating nowcasts

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

net impact = revision to nowcast

impacts of the news

revision to nowcast

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision


-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision

-0,3

-0,2

-0,1

0

0,1

0,2

0,3

0,4S

tart

ing p

oin

t: 1

7/0

7/2

01

5

NE

WS

1: 0

1/0

8/2

01

5

NE

WS

2.0

: 1

6/0

8/2

01

5

NE

WS

2.1

: 1

6/0

8/2

01

5

NE

WS

2.2

: 1

6/0

8/2

01

5

NE

WS

3: 0

1/0

9/2

01

5

NE

WS

4: 1

6/0

9/2

01

5

NE

WS

5: 0

1/1

0/2

01

5

NE

WS

6: 1

6/1

0/2

01

5

NE

WS

7: 0

1/1

1/2

01

5

NE

WS

8.0

: 1

5/1

1/2

01

5

NE

WS

8.1

: 1

5/1

1/2

01

5

NE

WS

8.2

: 1

5/1

1/2

01

5

EA GDP(0)

DE GDP(0)

EA Forecast GDP(0)


GDP BE (rev +1Q)

GDP BE (0)

Retail sales DE
















Revision


GDP flash

2015Q3 publication

22

Quantifying the role of the news

► When updating the nowcasts, the weight of the news depends on

● the quality of the indicator (i.e. the correlation of the factor with the innovations)

● its timeliness (variables that come first will typically have a bigger weight)

[w1𝑘,𝑡 , … , w5

𝑘,𝑡 ] = E[yk,t Iv+1′ ] 𝐸 𝐼𝑣+1 𝐼𝑣+1

′ −1

E[y𝑘,𝑡 ℱ𝑟𝑒𝑓𝑟𝑒𝑠ℎ𝑒𝑑 − E[y𝑘,𝑡 ℱ𝑜𝑙𝑑 =

𝑗=1

𝐽

w𝑗𝑘,𝑡 y(𝑖,𝑡)𝑗 − E y(𝑖,𝑡)𝑗 ℱ𝑜𝑙𝑑

timeliness

quality

► By multiplying the weights by the standard deviation of the news associated with each data release (instead of a given realization of the news), we obtain the standard impact of news :

● this measure allows us to compare the average informative content of the different indicators when the object of interest is quarterly growth…

● and hence, allows us to provide a ranking of indicators

23

Empirical results

► Survey data has a very large impace (see Basselier et al. 2014 for

further details)

24

Ranking for euro area GDP flash, based on the standard

impact in the benchmark case

► Cumulative impact of data releases over a whole semester

► Confirms the dominance of soft data, but industrial production in the

euro area and Germany should not be neglected

► Notice relevance of NBB survey for EA growth prediction

0,00

0,05

0,10

0,15

0,20

0,25

0,30

0,35

25

FORECASTING EVALUATION

Comparing the model by

Basselier, de Antonio Liedo and Langenus (2016)

with specifications that include a larger proportion of

Belgian indicators

26


1. Introduction










27

Overview of the models tested with JDemetra+

Name Model GDP indicators Size

Benchmark:

R2D2

Basselier, de Antonio and Langenus (2014),

focus on the euro area GDP

BE (flash, 120),

EA (flash, 120),

DE (flash)

35

R2D2 + BE

… + Disaggregate hard data for BE (VAT

returns, industrial production, temporal

employment)

BE (flash, 120,

1Y, final)

EA (flash, 120),

DE (flash)

60

R2D2 + BE Large … + Disaggregate business surveys data for

Belgium

BE (flash, 120,

1Y, final)

EA (flash, 120),

DE (flash)

80

• All models share a VAR(4) specification for the factors, which is suitable

to account for very heterogenous dynamics in the series

• Data from 2000, so focus on the recent sample

• R2D2 contains only key indicators followed by analysts in Bloomberg and

Forex Factory (it includes data from France, Italy, Spain and Germany)

• Heterogeneous loading structure following De Antonio Liedo (2013) and

Basselier et al (2014)

28

Using the model in real-time

► R2D2 model is being updated at the monthly basis

-1

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data ecb-60

29



-1

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data ecb-30

30



-1

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data ecb0

31

Simulating real-time forecasts with JDemetra+

► R2D2 model: Real time data sent to the ECB vs JD+ simulation (60 days

before the end of the quarter)

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data fh(-60) ecb-60

32


► R2D2: Real time data sent to the ECB vs JD+ simulation (30 days before

the end of the quarter)

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data fh(-30) ecb-30

33


► R2D2 : Real time data sent to the ECB vs JD+ simulation (0 days

before the end of the quarter)

-1

-0,8

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash

True data fh(0) ecb0

34


► R2D2 : Real time data sent to the ECB vs JD+ simulation (29 days after


-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash


35


► R2D2 : Real time data sent to the ECB vs JD+ simulation (29 days after


-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

BE Flash


over-estimation

of GDP (Flash)

36


► Successful simulation based on a stylized release

calendar

► Over-estimation episodes:

● Will the flash be revised upwards and prove the model right in

the end?

● Most likely hypothesis, particularly relevant for the most recent

period: predominance of euro area data in detriment of Belgian

specific indicators

► Research question: does the performance of the model

improve with the addition of Belgian specific information or it

does not matter due to the large degree of synchronization

accross the economies represented in the data?

37


► Large Model (60 variables)

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

01/1

0/2

01

2

01/1

2/2

01

2

01/0

2/2

01

3

01/0

4/2

01

3

01/0

6/2

01

3

01/0

8/2

01

3

01/1

0/2

01

3

01/1

2/2

01

3

01/0

2/2

01

4

01/0

4/2

01

4

01/0

6/2

01

4

01/0

8/2

01

4

01/1

0/2

01

4

01/1

2/2

01

4

01/0

2/2

01

5

01/0

4/2

01

5

01/0

6/2

01

5

01/0

8/2

01

5

01/1

0/2

01

5

01/1

2/2

01

5

01/0

2/2

01

6

01/0

4/2

01

6

01/0

6/2

01

6

01/0

8/2

01

6

01/1

0/2

01

6

01/1

2/2

01

6

01/0

2/2

01

7

01/0

4/2

01

7

01/0

6/2

01

7

01/0

8/2

01

7

01/1

0/2

01

7

01/1

2/2

01

7

01/0

2/2

01

8

01/0

4/2

01

8

01/0

6/2

01

8

01/0

8/2

01

8

01/1

0/2

01

8

01/1

2/2

01

8

01/0

2/2

01

9

01/0

4/2

01

9

BE Flash

True data fh(-30) ecb-30 revised

38


► Very Large Model (80 variables)

-0,6

-0,4

-0,2

0

0,2

0,4

0,6

0,8

1

01/1

0/2

01

2

01/1

2/2

01

2

01/0

2/2

01

3

01/0

4/2

01

3

01/0

6/2

01

3

01/0

8/2

01

3

01/1

0/2

01

3

01/1

2/2

01

3

01/0

2/2

01

4

01/0

4/2

01

4

01/0

6/2

01

4

01/0

8/2

01

4

01/1

0/2

01

4

01/1

2/2

01

4

01/0

2/2

01

5

01/0

4/2

01

5

01/0

6/2

01

5

01/0

8/2

01

5

01/1

0/2

01

5

01/1

2/2

01

5

01/0

2/2

01

6

01/0

4/2

01

6

01/0

6/2

01

6

01/0

8/2

01

6

01/1

0/2

01

6

01/1

2/2

01

6

01/0

2/2

01

7

01/0

4/2

01

7

01/0

6/2

01

7

01/0

8/2

01

7

01/1

0/2

01

7

01/1

2/2

01

7

01/0

2/2

01

8

01/0

4/2

01

8

01/0

6/2

01

8

01/0

8/2

01

8

01/1

0/2

01

8

01/1

2/2

01

8

01/0

2/2

01

9

01/0

4/2

01

9

Very Large Model: BE Flash

True data fh(-30) ecb-30 revised

39


► Increasing the number of variables seems to improve

performance over our small evaluation period

► Would the model have improved the performance over

the recession episode in spite of the large degree of

syncronization accross variables?

[TO BE DONE]

► Does the addition of the new Belgian indicators

change the ranking of most relevant indicators?

[TO BE DONE]

40

CASE STUDY

41


1. Introduction








4. Case study: Belgian and euro area GDP growth in 2019Q3


42

Using the model in practice

► At the end of august, the model expects -0.19% growth for Q3 in

Belgium. Today’s update suggests a slight improvement: -0.12%

43

The model “reads” all the data releases (43 in this

case) and precisely accounts for the 0.08pp revision

44

CONCLUSION

45


1. Introduction










46

Concluding remarks

► JDemetra+ is able to simulate real time forecasts that are very close to

the ones produced by the NBB economics department.

► Unclear whether enlarging the model improves performance, since the

evaluation sample considered is very small

► Since the model is successful at simulating real time forecast: increase

evaluation sample is the next thing to do

► FYI: you can reproduce the results by using JDemetra+ software1 and the

nowcasting plugin

1 JDemetra+ is free and open source software written in Java. Download it here: https://github.com/jdemetra/jdemetra-app/releases/tag/v2.1.0-rc2.

The Nowcasting plugin should be downloaded here: https://github.com/nbbrd/jdemetra-nowcasting/releases. After downloading it, go to the Tools option in

JDemetra+ and select plugins. The sofware is portable and it could even be executed from a USB disc.

https://github.com/jdemetra/jdemetra-app/releases/tag/v2.1.0-rc2

https://github.com/nbbrd/jdemetra-nowcasting/releases

A New Dynamic Factor Model for Nowcasting Belgian GDP growth · Introduction 2. The analysis of...

Documents

Transcript of A New Dynamic Factor Model for Nowcasting Belgian GDP growth · Introduction 2. The analysis of...