2-1 Chapter 2 Perception. 2-2 “Perception is Reality” Louis Cheskin.
Bubbles and Crashes: Perception vs. Reality...Perception vs. Reality National Bank of Belgium...
Transcript of Bubbles and Crashes: Perception vs. Reality...Perception vs. Reality National Bank of Belgium...
Bubbles and Crashes:Perception vs. Reality
National Bank of Belgium
Brussels, March 2019
William N. GoetzmannYale School of Management
Three Studies
• What Happens after a Boom?• What Happens after a Crash?• Probability Assessment of a Crash.
-
100
200
300
400
500
600
700
800
900
1.000
-
2.000
4.000
6.000
8.000
10.000
12.00026
-7-1
718
2-9-
1718
10-1
0-17
1817
-11-
1718
25-1
2-17
181-
2-17
1911
-3-1
719
18-4
-171
926
-5-1
719
3-7-
1719
10-8
-171
917
-9-1
719
25-1
0-17
192-
12-1
719
9-1-
1720
16-2
-172
025
-3-1
720
2-5-
1720
9-6-
1720
17-7
-172
024
-8-1
720
1-10
-172
08-
11-1
720
16-1
2-17
20
Mississippi Company Shares 1719 -1720
Mississippi Company
South Sea Company
US S&P Index to 1926-2018
I. How Common Are Bubbles?
• Bubble is a boom followed by a crash• Boom:
(A) >100% growth in 1 year(B) >100% growth over 3 years (like 1928 & 1999)
• Crash:(1) >50% decline the NEXT year(2) >50% decline after 5 years(3) > 50% decline sometime in the next 5 years
Data: World Stock Markets
• Dimson, Marsh and Staunton– Continuous, real annual returns for 21 countries
from 1900– Constructed from primary sources
• Jorion-Goetzmann (LofN, IFC, hand)– Discontinuous, monthly dollar-denominated
market appreciation– Constructed from (mostly) secondary sources
Frequency of booms and busts conditional on astock market index increasing by 100% ordecreasing by 50% in a single calendar year
(A) (B) (C) (D) (E)
Fullsample
Nextyearboom
Nextyearcrash
Five yearboom
Five year crash
(+100%) (-50%) (+100%) (-50%)
One-yearboom
58 4 4 13 101.7% 6.9% 6.9% 22.4% 17.2%
One-yearcrash
67 9 1 22 21.9% 13.4% 1.5% 32.8% 3.0%
All market-years
3514 56 74 592 2981 1.6% 2.1% 16.9% 8.5%
Frequency of booms and busts conditional on astock market index increasing by 100% ordecreasing by 50% over a 3-year calendar period
(A) (B) (C) (D) (E)
Fullsample
Nextyearboom
Nextyearcrash
Five yearboom
Five yearcrash
(+100%) (-50%) (+100%) (-50%)
Three-yearboom
346 12 16 57 3410.1% 3.5% 4.6% 16.5% 9.8%
Three-yearcrash
202 12 12 68 235.9% 5.9% 5.9% 33.7% 11.4%
All market-years3412 54 68 587 287100% 1.6% 2.0% 17.2% 8.4%
0,0
0,5
1,0
1,5
2,0
2,5
3,0
3,5
t-5 t-4 t-3 t-2 t-1 t0 t+1 t+2 t+3 t+4 t+5
Markets with a greater than 100% return in agiven year (all data)
41 Markets3,441 market years70 with a 100% boom
11 of 70 of those decline by 50% in five years3 of 70 decline by 50% in one year
0
0,5
1
1,5
2
2,5
3
3,5
4
4,5
t-5 t-4 t-3 t-2 t-1 t = 0 t+1 t+2 t+3 t+4 t+5
Event study: 41 markets, 433 eventsBubble defined as doubling in 3 years
mean
median
5%
95%
25%
75%
Reverse BubblesResearch with Co-author Dasol Kim
• Investors overly pessimistic or fearful?• May leave the market• May lose confidence in the market
II. Crashes
Frequency of booms and busts conditional on astock market index increasing by 100% ordecreasing by 50% in a single calendar year
(A) (B) (C) (D) (E)
Fullsample
Nextyearboom
Nextyearcrash
Five yearboom
Five year crash
(+100%) (-50%) (+100%) (-50%)
One-yearboom
58 4 4 13 101.7% 6.9% 6.9% 22.4% 17.2%
One-yearcrash
67 9 1 22 21.9% 13.4% 1.5% 32.8% 3.0%
All market-years
3514 56 74 592 2981 1.6% 2.1% 16.9% 8.5%
Frequency of booms and busts conditional on astock market index increasing by 100% ordecreasing by 50% over a 3-year calendar period
(A) (B) (C) (D) (E)
Fullsample
Nextyearboom
Nextyearcrash
Five yearboom
Five yearcrash
(+100%) (-50%) (+100%) (-50%)
Three-yearboom
346 12 16 57 3410.1% 3.5% 4.6% 16.5% 9.8%
Three-yearcrash
202 12 12 68 235.9% 5.9% 5.9% 33.7% 11.4%
All market-years3412 54 68 587 287100% 1.6% 2.0% 17.2% 8.4%
Global Financial Data [GFD]
• Monthly returns 100+ countries• 1695-present• Conditional distributions: x % drop• Subsequent one-year return• Control for:
– Financial shocks– Macroeconomic shocks– Wars
-10%
-5%
0%
5%
10%
15%
20%
negative -10% -20% -30% -40% -50%
Market Decline in One Year
Next Year Market ReturnGFD data from 1690-2016
100+ markets
Controlling For:
• Institutional frictions• Financial crises• Macroeconomic shocks• Political conflicts• Survivorship issues
III. Behavior
Affect, Media and EarthquakesWilliam Goetzmann, Dasol Kim
and Robert Shiller
OverviewDirectly observe investor crash beliefs over time and incross section.
Test whether these are:Consistent with historical distributionInfluenced by availability heuristicInfluenced by media attention.
Rely on cross-section of investor types and onexogenous priming shocks about rare events.Test whether beliefs matter to investor choice.
Opinion Data: Shiller StockMarket Confidence Surveys
Mail survey from 1989-Present.
Monthly since 2001.
Individual and institutional investors.
9,953 responses (1989-2015)
Location (5-digit ZIP) from 2007.
The percent of the sample who attach little probabilityto a stock market crash in the next-six months.
Crash Question:
“What do you think is the probability of acatastrophic stock market crash in the U. S., likethat of October 28, 1929 or October 19, 1987, in
the next six months, including the case that acrash occurred in the other countries and spreads
to the U. S.? (An answer of 0% means that itcannot happen, an answer of 100% means it is
sure to happen.)
Probability in U. S.:_______________%”
0%
5%
10%
15%
20%
25%
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015
Wor
stda
ilyPe
rcen
tage
Dro
pin
DJI
Ain
the
Year
Subj
ectiv
ePr
obab
ility
ofC
rash
&An
nual
ized
Vola
tility
annualized volatility minimum daily return in the yearInstitutional 6 month crash probability Individual 6 month crash probability
Crash Probabilities Versus Market Volatility:1989 - 2015
High Crash Probability
10% to 20% vs. historical USfrequency of < 2%.
Covaries with implied and actualvolatility.
Differences between individual,institutional investors.
Affect Heuristic
“The results … demonstrate that moodinduced by brief reports has a large and
pervasive impact on estimates of thefrequency of risks and other undesirable
events. Furthermore, the effect isindependent of the similarity between
the story and the risk.”
(Johnson & Tversky, 1983)
Natural Experiments• Exogenous shock unrelated to future
crash probability.• Priming negative or positive effect on
mood.• Test for differential crash probability• Difference between individual and
institutional investors.
Earthquake
Priming mechanism: recent, moderate earthquakes(2.5- 5.5) within 30 miles of investor location.
Controlling for severe earthquakes (5.5+),earthquake frequency, baseline control variables.
Bootstrapped p-values.
Prediction: people more prone to availability & affectheuristics will report higher crash probability.
• Feeling frightened, on edge, nervous, and tense.• Being easily startled and looking out for danger.• Feeling various emotions such as fear, sadness, grief, guilt.• Anxiety symptoms like a racing heart, rapid breathing,trembling.• Impaired concentration, decision making and memory.• Worrying about what might have been or having to deal withreal ongoing concerns.• Feeling a sense of lack of control.• Thoughts and memories about the event continuing to popinto your mind, even days or weeks afterwards.• Feeling like the distressing events are happening again (i.e.,flashbacks)…
Earthquake Anxiety
Investor Subsample: Indiv. Indiv. Inst. Inst.
Dependent Variable: CrashProbability
CrashProbability
CrashProbability
CrashProbability
ModerateEarthquake (t-30,t) 0.053** 0.051** -0.015 -0.016
Quake FrequencyControl VariablesDay of Week FEMonth FE
YesNoNoNo
YesYesYesYes
YesNoNoNo
Yes,YesYesYes
NR-square
3,0960.29%
3,0961.32%
3,4880.03%
3,4881.16%
Crash Probabilities & Nearby Earthquakes
Lotteries• Lottery ticket sales and lottery stocks
– Kumar, Page & Spalt (2016)– Chen, Kumar & Zheng (2018)
• Proximity to winners increases consumptionand bankruptcy
• Kuhn et. al., 2011;• Agarawal et. al. 2018
Investor Subsample: Indiv. Indiv. Inst. Inst.
Dependent Variable: CrashProbability
CrashProbability
CrashProbability
CrashProbability
Nearby LotteryWinner (t-30,t) -0.072** -0.071** 0.038 0.038
Control VariablesDay of Week FEMonth FE
NoNoNo
YesYesYes
NoNoNo
YesYesYes
NR-square
2,7930.85%
2,7931.61%
3,1320.65%
3,1321.24%
Crash Probabilities & Nearby Winning Lottery Ticket Sale
The key explanatory variables are dummies associated with whether the investor islocated within a 30-mile radius of the store that sold a winning lottery ticket associatedwith Powerball or MegaMillions contests within the past 30 days
-0,08
-0,06
-0,04
-0,02
0
0,02
0,04
0,06
Earthquake Lottery
Conditional Crash Probabilities
Individual Institutional
Media Role
• Daily news articles about stock market• Calculate sentiment score• Use as predictor of individual and institutional crash
probabilities• Include controls:
• returns, past returns, past negative/positive articles,past sentiment, past returns volatility, past averagecrash probabilities, number of news articles,weekday/month FE.
• Include sentiment of first paragraph controlling for entirearticle sentiment.
Sources and MethodsProQuest Wall Street Journal search
Stock market articles: i.e., (stock NEAR/5 market) ORSU(stock) OR SU(securities).
133,496 articles with a word count of at least 250 wordsGeneral Inquirer: Negative (2,291 terms) and Positive
(1,915 terms) valence word lists.Sentiment measures (daily)
TF-IDF Weighting: Loughran and Macdonald (2011) forrelative frequency and article length.
“Positive articles” = # articles in top 10th samplepercentile for article-level sentiment.
“Negative articles” = # articles in bottom 10th.“Sentiment” = (௦௧௩ ି ௧௩ ௪ௗ௦)
(௦௧௩ ା ௧௩ ௪ௗ௦)across all articles
Table 6: Crash Probabilities and Article Counts
The table displays the results from OLS regression models where the dependent variables are the investor crashprobabilities (p). The results are displayed separately for individual (Indiv) and institutional (Inst) investors, and wherethe return variables are based upon the CRSP-VW (NYSE/AMEX/Nasdaq/Arca), S&P 500 or DJIA indices. Standarderrors are clustered on the date level and are displayed in parentheses. Statistical significance at the 1%, 5% and 10%levels are denoted as ***, **, and *, respectively.
(1) (2) (3) (4) (5) (6)Investor Subsample: Indiv. Inst. Indiv. Inst. Indiv. Inst.Return Variable: All All S&P500 S&P500 DJIA DJIADependent Variable: p (i,t) p (i,t) p (i,t) p (i,t) p (i,t) p (i,t)
r (t-1) × Negative (t) -1.012*** -0.168 -0.939*** -0.152 -1.040*** -0.212(0.363) (0.295) (0.365) (0.276) (0.394) (0.283)
r (t-1) × Positive (t) 0.206 0.239 0.333 0.189 0.476 0.310(0.429) (0.356) (0.432) (0.343) (0.462) (0.364)
Negative (t) 0.006 0.009** 0.007 0.009** 0.007 0.009**(0.006) (0.004) (0.006) (0.004) (0.006) (0.004)
Positive (t) -0.001 0.001 -0.001 0.001 -0.002 0.001(0.005) (0.005) (0.005) (0.005) (0.006) (0.004)
Sentiment (t-30,t-1) 0.036 -0.024 0.037 -0.021 0.028 -0.034(0.126) (0.098) (0.128) (0.099) (0.128) (0.102)
r (t-1) 0.068 -0.204 0.121 -0.186 0.136 -0.219(0.262) (0.217) (0.258) (0.212) (0.278) (0.221)
r (t-30,t-2) -0.174*** -0.036 -0.179*** -0.031 -0.195*** 0.006(0.061) (0.052) (0.064) (0.055) (0.065) (0.056)
s (t-30,t-2) 1.725*** 0.951** 1.627*** 1.006** 1.505*** 1.245**(0.548) (0.485) (0.546) (0.486) (0.581) (0.527)
p (t-30,t-1) 0.223*** 0.251*** 0.228*** 0.251*** 0.242*** 0.250***(0.062) (0.056) (0.061) (0.056) (0.061) (0.056)
log(1+#Articles (t)) 0.019*** 0.001 0.019*** 0.001 0.019*** 0.001(0.007) (0.007) (0.007) (0.007) (0.007) (0.007)
Weekday FEs YES YES YES YES YES YESMonth FEs YES YES YES YES YES YES
N 4730 6253 4730 6253 4730 6253Adjusted R2 1.99% 1.22% 1.96% 1.21% 1.92% 1.23%
Investor Subsample: Individual Institutional
Dependent Variable: CrashProbability
CrashProbability
Interaction of NegativeArticles (t) & Return (t-1) -0.939*** -0.152
Interaction of PositiveArticles (t) & Return (t-1) 0.333 0.343
• Control variables: returns, past returns, pastnegative/positive articles, past sentiment, past returnsvolatility, past average crash probabilities, number ofnews articles, weekday/month FEs.
Crash Probabilities:Negative Versus Positive Article Counts (S&P 500)
Table 7: Crash Probabilities and Media Sentiment
The table displays the results from OLS regression models where the dependent variables are the investor crashprobabilities (p). The results are displayed separately for individual (Indiv.) and institutional (Inst.) investors, andwhere the return variables are based upon the CRSP-VW (NYSE/AMEX/Nasdaq/Arca), S&P 500 or DJIA indices.Standard errors are clustered on the date level and are displayed in parentheses. Statistical significance at the 1%, 5%and 10% levels are denoted as ***, **, and *, respectively.
(1) (2) (3) (4) (5) (6)Investor Subsample: Indiv. Inst. Indiv. Inst. Indiv. Inst.Return variable: All All S&P500 S&P500 DJIA DJIADependent Variable: p (i,t) p (i,t) p (i,t) p (i,t) p (i,t) p (i,t)
r (t-1) × Sentiment (t) 6.881* 0.482 7.223* -0.344 8.832** 0.102(3.932) (3.440) (3.925) (3.132) (4.267) (3.396)
Sentiment (t) -0.036 -0.053 -0.038 -0.055 -0.039 -0.056(0.053) (0.041) (0.053) (0.042) (0.053) (0.041)
Sentiment (t-30,t-1) 0.014 -0.017 0.018 -0.015 0.009 -0.025(0.131) (0.100) (0.131) (0.100) (0.131) (0.104)
r (t-1) -0.683* -0.336 -0.667* -0.260 -0.820** -0.358(0.363) (0.280) (0.359) (0.257) (0.390) (0.271)
r (t-30,t-2) -0.163*** -0.038 -0.172*** -0.034 -0.191*** 0.001(0.062) (0.052) (0.065) (0.055) (0.066) (0.062)
s (t-30,t-2) 1.605*** 0.972** 1.522*** 1.008** 1.380** 1.253**(0.552) (0.493) (0.548) (0.494) (0.580) (0.535)
p (t-30,t-1) 0.228*** 0.246*** 0.235*** 0.247*** 0.250*** 0.246***(0.061) (0.057) (0.061) (0.057) (0.061) (0.058)
log(1+#Articles (t)) 0.022*** 0.009 0.023*** 0.009 0.022*** 0.009*(0.006) (0.005) (0.006) (0.005) (0.006) (0.006)
Weekday FEs YES YES YES YES YES YESMonth FEs YES YES YES YES YES YES
N 4730 6253 4730 6253 4730 6253Adjusted R2 1.87% 1.20% 1.91% 1.20% 1.92% 1.20%
Investor Subsample: Individual Institutional
Dependent Variable: CrashProbability
CrashProbability
Interaction of Sentiment(t) & Return (t-1) [S&P500]
7.223* -0.344
Interaction of Sentiment(t) & Return (t-1) [DJIA] 8.832** 0.102
• Control variables (Table 6)
Crash Probabilities:Media Sentiment Measure
Table 9: Salience of Lead Paragraph and Article Placement
The table displays the results from OLS regression models where the dependent variables are the investor crashprobabilities (p). The results are displayed separately for individual (Indiv) and institutional (Inst) investors, and wherethe return variables are based upon the CRSP-VW (NYSE/AMEX/Nasdaq/Arca), S&P 500 or DJIA indices. Controlvariables of Table 7 are included in all the models but not all are reported. Standard errors are clustered on the datelevel and are displayed in parentheses. Statistical significance at the 1%, 5% and 10% levels are denoted as ***, **,and *, respectively.
Panel A: Lead Paragraph
(1) (2) (3) (4) (5) (6)Investor Subsample: Indiv. Inst. Indiv. Inst. Indiv. Inst.Return Variable: All All S&P500 S&P500 DJIA DJIADependent Variable: p (i,t) p (i,t) p (i,t) p (i,t) p (i,t) p (i,t)
r (t-1) × SentimentLead (t) 4.451** 0.093 4.545** -0.017 5.726** 0.150(2.130) (1.556) (2.206) (1.737) (2.357) (1.668)
r (t-1) × SentimentNotLead (t) 1.366 -0.696 1.440 -1.276 2.520 -0.688(3.254) (2.576) (3.200) (2.408) (3.360) (2.648)
SentimentLead (t) -0.037 -0.027 -0.039 -0.027 -0.042 -0.026(0.030) (0.022) (0.030) (0.022) (0.030) (0.022)
SentimentNotLead (t) 0.070** 0.023 0.069** 0.022 0.069** 0.021(0.032) (0.027) (0.032) (0.027) (0.032) (0.026)
Control Variables YES YES YES YES YES YESWeekday FEs YES YES YES YES YES YESMonth FEs YES YES YES YES YES YES
N 4730 6253 4730 6253 4730 6253Adjusted R2 2.01% 1.18% 2.03% 1.18% 2.04% 1.19%
Investor Subsample: Individual Institutional
Dependent Variable: CrashProbability
CrashProbability
Interaction of LeadParagraph Sentiment (t)& Return (t-1)
4.545** -0.017
Crash Probabilities:Lead Paragraph Sentiment (S&P 500)
• Control variables (Table 6); non-lead paragraphsentiment (t); interaction of non-lead paragraph (t) andreturn (t-1).
Table 9: Salience of Lead Paragraph and Article Placement
The table displays the results from OLS regression models where the dependent variables are the investor crashprobabilities (p). The results are displayed separately for individual (Indiv) and institutional (Inst) investors, and wherethe return variables are based upon the CRSP-VW (NYSE/AMEX/Nasdaq/Arca), S&P 500 or DJIA indices. Controlvariables of Table 7 are included in all the models but not all are reported. Standard errors are clustered on the datelevel and are displayed in parentheses. Statistical significance at the 1%, 5% and 10% levels are denoted as ***, **,and *, respectively.
Panel B: Article Placement
(1) (2) (3) (4) (5) (6)Investor Subsample: Indiv. Inst. Indiv. Inst. Indiv. Inst.Return Variable: All All S&P500 S&P500 DJIA DJIADependent Variable: p (i,t) p (i,t) p (i,t) p (i,t) p (i,t) p (i,t)
r (t-1) × SentimentFront (t) 4.485* 2.810 4.952** 2.533 5.099* 3.006(2.373) (1.993) (2.381) (1.963) (2.670) (2.194)
r (t-1) × SentimentNotFront (t) 2.174 -1.835 2.152 -2.172 3.330 -1.725(3.295) (3.059) (3.310) (2.896) (3.659) (2.975)
SentimentFront (t) -0.012 -0.016 -0.012 -0.016 -0.013 -0.017(0.029) (0.024) (0.028) (0.024) (0.029) (0.024)
SentimentNotFront (t) -0.049 -0.046 -0.052 -0.047 -0.051 -0.048(0.043) (0.035) (0.043) (0.035) (0.044) (0.035)
Control Variables YES YES YES YES YES YESWeekday FEs YES YES YES YES YES YESMonth FEs YES YES YES YES YES YES
N 4730 6253 4730 6253 4730 6253Adjusted R2 1.87% 1.21% 1.92% 1.20% 1.93% 1.21%
Investor Subsample: Individual Institutional
Dependent Variable: CrashProbability
CrashProbability
Interaction of FrontPage PlacementSentiment (t) & Return(t-1)
4.952** 2.533
Front Page Article Sentiment (S&P 500)
• Control variables (Table 6); non-front page sentiment(t); interaction of non-front page sentiment (t) andreturn (t-1).
Investor Response to Media &Market
We show that the negative media sentiment issignificantly associated with investor crash
beliefs, but only for individual investors.If results driven by fundamental factors, effectsshould be similar for both investor types. This is
not the case.
Consistent with affect heuristic moderated byprofessional experience or analytical
framework.
Consistent with investor sophisticationinfluencing belief expectations.
Investor Choice TestsDaily Fund Flow Data: TrimTabs
2003-2015 (greater frequency of survey data)Equity Fund Net FlowsFixed Income Fund Net Flows
Rolling weekly probabilities: individual andinstitutional
Result: Individual investor crash probabilitiessignificantly lead equity mutual fund outflows, but donot for institutional investors.
-0,14
-0,12
-0,1
-0,08
-0,06
-0,04
-0,02
0
0,02
0,04
t+1 t+2 t+3 t+4 t+5 t+6 t+7 t+8 t+9 t+10
upper pIndiv. (t) lower
Aggregate Equity Daily Fund Flows:Individual Investor Crash Probability Coefficients
-0,08
-0,06
-0,04
-0,02
0
0,02
0,04
0,06
0,08
0,1
t+1 t+2 t+3 t+4 t+5 t+6 t+7 t+8 t+9 t+10
upper pIndiv. (t) lower
Aggregate Equity Daily Fund Flows:Institutional Investor Crash Probability Coefficients
Summary• Bubbles are rare• Market tends to rebound after serious crash• High probability of a market crash, consistent with
equity risk premium.• Natural experiments consistent with affect heuristic
influencing belief formation.• Media plays a negative feedback role for individual
investor beliefs.• Crash beliefs have meaningful effects on realized
investor outcomes (i.e., fund flows).