Crash Narratives William Goetzmann (Yale), Dasol
Description: Crash Narratives William Goetzmann (Yale), Dasol Kim (OFR), Bob Shiller (Yale) Financial Stability Conference November 17, 2022 The opinions and views of the authors do not necessarily reflect the positions or policies of the Office of
Related Topics
Download Presentation
"Crash Narratives William Goetzmann (Yale), Dasol" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. Crash Narratives William Goetzmann (Yale), Dasol Kim (OFR), Bob Shiller (Yale)
Financial Stability Conference
November 17, 2022
The opinions and views of the authors do not necessarily reflect the positions or policies of the Office of Financial Research or the US Department of the Treasury.<br>
slide2. “When Land Becomes Water…”<br>
slide4. Narratives in Collective Memory Few people experience extreme market crashes in their lifetimes, and most often rely on collective rather than personal memory.
Collective memory is usually represented by narratives rather than comprehensive accounts.
Narratives generally elicit causal relationships, providing analogies for how and why things happen, and what the future might be.
Financial press is an important conduit.
Journalists may use narratives to contextualize current events.
Choice of narratives is subjective, and may not be necessarily congruent with actual conditions.
Investors may rely on media narratives to form beliefs and inform choices, implying aggregate market feedback (Shiller, 2017).
This paper examines the role of financial media in propagating narratives that may go viral.<br>
slide5. Overview Develop a measure of media narratives using articles published following major market crashes, i.e., crash narratives.
Strong association between crash narratives and traditional market- as well as survey-based crash indicators.
Evidence of virality of media narratives, feedback on market volatility.
Investors revise crash assessments following the appearance of crash narratives in media, and parrot the same language in written responses.
Media increases general or pure narrativity during periods of high investor attention. General narrativity drives main results.<br>
slide6. Related Literature Financial and economic narratives.
Bybee et al. (2021), Bertesch et. al. (2021), Larsen et. al. (2021), Dierckx et al. (2021).
Rare disasters.
Reitz (1988), Barro (2006), Weitzmann (2007), Santa-Clara & Yan (2010), Berkman et. al. (2011), Bollerslev & Todorov (2011), Wachter (2013), Welch (2015), Goetzmann, Kim & Shiller (2022).
Memory models.
Mullainathan (2002), Gennaioli and Shleifer (2010), Kahana (2012), Bordalo et. al. (2017), Enke et. al. (2019), Wachter and Kahana (2019), Goetzmann, Watanabe & Watanabe (2022).<br>
slide7. What is a Narrative? Economic narratives are contagious stories with the potential to change how people make economic decisions independent of fundamentals (Shiller, 2019).
Narrativity is a set of qualities or properties that distinguishes narratives from non-narrative texts (Herman, 2009):
Situatedness (voice indicating occasion for telling)
Relevance of event sequencing (causality)
Worldmaking, disruption (disequilibrium)
What it’s like (experience of story-world)
Quantifying narratives (or narrativity) is intrinsically difficult as it involves high-dimensional semantic representations.<br>
slide8. Article from September 16, 2008:
If you learn nothing else from the last few harrowing days, you should learn the difference between what is obvious and what is inevitable.
In the heat of the moment, the two perceptions seem identical. It was obvious that investors would panic as they absorbed the news about Bloody Sunday on Wall Street, so it was inevitable that the market would take a slashing. It was obvious that Lehman Brothers had to go bust, so a bankruptcy filing was inevitable. It was obvious that Merrill Lynch could no longer make it on its own, so it was inevitable that a bigger institution like Bank of America would take it over.
But investors -- at least individual investors -- don't actually panic in times like these. Instead, they freeze. In July (the latest month for which final numbers are available), mutual-fund investors pulled out just $2.62 of every $100 they had invested in stock funds. That was less than they took out of bond funds, even though the stock market had just gone through a nauseating summer swoon.
[...] Example: September 16, 2008<br>
slide9. Quantifying “Crash Narratives” Recent advances in computational linguistics potentially provide a way of quantifying narratives.
Distributional Hypothesis: Linguistic terms with similar distributions (used or occur in same contexts) can purport similar meanings.
Statistical semantics builds on this to provide mathematical representations of ideas based on sequencing of words, sentences, etc.
If we can properly model the distributed semantic representations (for subjective accounts of stock market crashes), it may be possible to identify the narrativity (crash narratives) broadly.<br>
slide10. Quantifying “Crash Narratives” (cont.) Our main approach is based on established distributional semantic methods that allows us compare semantic similarity across articles over time.
Doc2Vec (Le & Mikolov, 2014) is based on neural network modeling to create distributed semantic representations of words, paragraphs, etc.
Evidence of effectiveness for classifying narratives (Dehghani et. al., 2017).
Complementary to other approaches (e.g., BERT)
We focus on periods rich in crash narratives: periods immediately following stock market crashes (e.g., Black Monday).
Use Doc2Vec to quantify semantic structures and compare the similarity in those structures across periods.
Evolving lexicon of financial press standardization
‘87 Narrative =
(Cosine Similarity with Articles on Oct. 20 - 23, 1987)
- (Cosine Similarity with Articles on Oct. 5 - 9, 1987)<br>
slide11. Example: September 16, 2008 (cont.) BEST matched article from Oct 20, 1987 to Sept 16, 2008 example:
The little investor doesn't know where to turn, although bigger ones are putting up a brave front.
"I'm scared," says Julie Ianotti, an executive secretary in Houston. "My stock is my nest egg, for a house or something. Should I sell? Tell me, should I sell?"
When Ms. Ianotti's mother suggested the price drop could portend a new depression, the young Ms. Ianotti at first scoffed at the idea. "I said, 'Yeah, sure, Mom. It'll go back up.'" But as the Dow Jones Industrial Average was taking a record plunge yesterday, Ms. Ianotti says she began giving more thought to her mother's stories about subsisting on water and sugar in the Depression.
[...]<br>
slide12. WORST matched article from Oct 20, 1987 to Sept 6, 2008 example:
IC Industries Inc., expanding a restructuring of the company, said it will buy as much as $1 billion of its common shares and is considering selling its aerospace unit, valued at more than $1.5 billion.
Karl D. Bays, chairman and chief executive officer, said the moves are being undertaken to focus the company around its most profitable businesses, which include food products, soft-drink bottling and auto repair. He said the steps aren't related to recent market reports that Minneapolis investors Irwin Jacobs and Carl Pohlad may seek control of the company.
In September, IC announced plans to divest its Illinois Central Gulf Railroad unit by distributing shares in the division to IC holders. It estimated the railroad's market value at $250 million.
[...] Example: September 16, 2008 (cont.)<br>
slide13. Prevalence of Crash Narratives<br>
slide14. Data Financial News data:
ProQuest Wall Street Journal Eastern Edition (1984 – 2020).
231,230 articles.
Hand-collected Wall Street Journal data from 1929.
Investor survey data:
Investor Behavior Project (Shiller Survey).
19,426 responses (1989 – 2020).
Crash probability estimates, general comments, date of response, investor type (i.e. individual, institutional).
Internet search data:
Google Trends API.<br>
slide15. 1. How Do Narratives Propagate? Assess virality using internet search volumes (Da et al., 2011).
Investor interest related to crashes (CrashAttention) should be higher after seeing crash narratives in financial press.
Effects should be conditioned on general stock market attention.
Aggregate search volumes, not user-level.
Model specification:
CrashAttentiont+1 = β1 × ’87 Narrativet + β2 × StockMarketAttentiont + β3 × ’87 Narrativet × StockMarketAttentiont + Σj=15 β4,j × Xt-j+1 + Σj=15 β5,j × CrashAttentiont-j+1 + κDOWt + κMontht + ηt+1
Control variables (X): VIX, option-implied crash probabilities, market returns, market volatility, turnover.<br>
slide16. 1. Result: Propagation Evidence of positive feedback loop related to media narratives.
Other results:
Weaker effects for crash narratives from ‘29 crash (salience).
Crash narratives have positive feedback on market volatility.<br>
slide17. 2. Do Narratives Affect Investor Beliefs? Fundamental news captured by the narrative measure will likely affect individual and institutional investors similarly.
However, individuals may rely more on judgement heuristics when expertise low (Ottati & Isbell, 1996; Isbell & Wyer, 1999).
Focus on differential effect based on investor sophistication.
Model specification (by investor type):
πAdji,t = γ1 × ’87 Narrativet-1 + γ2 × StockMarketAttentiont-1 + γ3 × ’87 Narrativet-1 × StockMarketAttentiont-1 + γ4 × Wi,t-1 + λDOWt + λYear-Montht + ζi,t
Adjusted crash probabilities (πAdj): survey crash probability minus option-implied crash probability.
Control variables (W): VIX, market returns, past month market returns, market volatility, past month survey probabilities, past month stock market attention, past month crash narratives.<br>
slide18. 2. Result: Investor Crash Beliefs Individual investors adjust crash probabilities in response to crash narratives, but not institutional investors.
Other results:
Bag-of-words measures have little impact on investor assessments.
Investors “parrot” crash narratives in general assessments.<br>
slide19. 3. When Does Media Use Narratives? Pure narrativity that is plausibly exogenous to the financial markets.
Brothers Grimm Fairy Tales widely used as archetypal narrative corpora.
Stories by motif category, estimate similarity with each WSJ article.
Use the first principal component factor to capture general narrativity.
We find evidence (not shown) that journalists use narratives to make information more accessible during periods of broader readership.
Allows us to directly assess whether the crash narrative results are actually due to narrativity rather than other factors.<br>
slide20. 3. Pure Narrativity (FolkMotif) High correlation between crash and pure narratives measures: ~90%<br>
slide21. 3. Result: Narrativity, Investor Beliefs Effects isolated to component related to “pure” narratives, confirming that the main results are related to narrativity.<br>
slide22. Closing History plays an important role in press narratives by bringing attention to past “rare” events in U.S. stock market history.
Our results suggest investors rely on narrative heuristics in forming beliefs, affecting aggregate market outcomes.
Collective and individual memory processes are not mutually exclusive.
Stories may also be a mechanism to perpetuate a particularly important event in individual memory.
Interactions between collective and individual memory processes are not well understood.<br>
Financial Stability Conference
November 17, 2022
The opinions and views of the authors do not necessarily reflect the positions or policies of the Office of Financial Research or the US Department of the Treasury.<br>
slide2. “When Land Becomes Water…”<br>
slide4. Narratives in Collective Memory Few people experience extreme market crashes in their lifetimes, and most often rely on collective rather than personal memory.
Collective memory is usually represented by narratives rather than comprehensive accounts.
Narratives generally elicit causal relationships, providing analogies for how and why things happen, and what the future might be.
Financial press is an important conduit.
Journalists may use narratives to contextualize current events.
Choice of narratives is subjective, and may not be necessarily congruent with actual conditions.
Investors may rely on media narratives to form beliefs and inform choices, implying aggregate market feedback (Shiller, 2017).
This paper examines the role of financial media in propagating narratives that may go viral.<br>
slide5. Overview Develop a measure of media narratives using articles published following major market crashes, i.e., crash narratives.
Strong association between crash narratives and traditional market- as well as survey-based crash indicators.
Evidence of virality of media narratives, feedback on market volatility.
Investors revise crash assessments following the appearance of crash narratives in media, and parrot the same language in written responses.
Media increases general or pure narrativity during periods of high investor attention. General narrativity drives main results.<br>
slide6. Related Literature Financial and economic narratives.
Bybee et al. (2021), Bertesch et. al. (2021), Larsen et. al. (2021), Dierckx et al. (2021).
Rare disasters.
Reitz (1988), Barro (2006), Weitzmann (2007), Santa-Clara & Yan (2010), Berkman et. al. (2011), Bollerslev & Todorov (2011), Wachter (2013), Welch (2015), Goetzmann, Kim & Shiller (2022).
Memory models.
Mullainathan (2002), Gennaioli and Shleifer (2010), Kahana (2012), Bordalo et. al. (2017), Enke et. al. (2019), Wachter and Kahana (2019), Goetzmann, Watanabe & Watanabe (2022).<br>
slide7. What is a Narrative? Economic narratives are contagious stories with the potential to change how people make economic decisions independent of fundamentals (Shiller, 2019).
Narrativity is a set of qualities or properties that distinguishes narratives from non-narrative texts (Herman, 2009):
Situatedness (voice indicating occasion for telling)
Relevance of event sequencing (causality)
Worldmaking, disruption (disequilibrium)
What it’s like (experience of story-world)
Quantifying narratives (or narrativity) is intrinsically difficult as it involves high-dimensional semantic representations.<br>
slide8. Article from September 16, 2008:
If you learn nothing else from the last few harrowing days, you should learn the difference between what is obvious and what is inevitable.
In the heat of the moment, the two perceptions seem identical. It was obvious that investors would panic as they absorbed the news about Bloody Sunday on Wall Street, so it was inevitable that the market would take a slashing. It was obvious that Lehman Brothers had to go bust, so a bankruptcy filing was inevitable. It was obvious that Merrill Lynch could no longer make it on its own, so it was inevitable that a bigger institution like Bank of America would take it over.
But investors -- at least individual investors -- don't actually panic in times like these. Instead, they freeze. In July (the latest month for which final numbers are available), mutual-fund investors pulled out just $2.62 of every $100 they had invested in stock funds. That was less than they took out of bond funds, even though the stock market had just gone through a nauseating summer swoon.
[...] Example: September 16, 2008<br>
slide9. Quantifying “Crash Narratives” Recent advances in computational linguistics potentially provide a way of quantifying narratives.
Distributional Hypothesis: Linguistic terms with similar distributions (used or occur in same contexts) can purport similar meanings.
Statistical semantics builds on this to provide mathematical representations of ideas based on sequencing of words, sentences, etc.
If we can properly model the distributed semantic representations (for subjective accounts of stock market crashes), it may be possible to identify the narrativity (crash narratives) broadly.<br>
slide10. Quantifying “Crash Narratives” (cont.) Our main approach is based on established distributional semantic methods that allows us compare semantic similarity across articles over time.
Doc2Vec (Le & Mikolov, 2014) is based on neural network modeling to create distributed semantic representations of words, paragraphs, etc.
Evidence of effectiveness for classifying narratives (Dehghani et. al., 2017).
Complementary to other approaches (e.g., BERT)
We focus on periods rich in crash narratives: periods immediately following stock market crashes (e.g., Black Monday).
Use Doc2Vec to quantify semantic structures and compare the similarity in those structures across periods.
Evolving lexicon of financial press standardization
‘87 Narrative =
(Cosine Similarity with Articles on Oct. 20 - 23, 1987)
- (Cosine Similarity with Articles on Oct. 5 - 9, 1987)<br>
slide11. Example: September 16, 2008 (cont.) BEST matched article from Oct 20, 1987 to Sept 16, 2008 example:
The little investor doesn't know where to turn, although bigger ones are putting up a brave front.
"I'm scared," says Julie Ianotti, an executive secretary in Houston. "My stock is my nest egg, for a house or something. Should I sell? Tell me, should I sell?"
When Ms. Ianotti's mother suggested the price drop could portend a new depression, the young Ms. Ianotti at first scoffed at the idea. "I said, 'Yeah, sure, Mom. It'll go back up.'" But as the Dow Jones Industrial Average was taking a record plunge yesterday, Ms. Ianotti says she began giving more thought to her mother's stories about subsisting on water and sugar in the Depression.
[...]<br>
slide12. WORST matched article from Oct 20, 1987 to Sept 6, 2008 example:
IC Industries Inc., expanding a restructuring of the company, said it will buy as much as $1 billion of its common shares and is considering selling its aerospace unit, valued at more than $1.5 billion.
Karl D. Bays, chairman and chief executive officer, said the moves are being undertaken to focus the company around its most profitable businesses, which include food products, soft-drink bottling and auto repair. He said the steps aren't related to recent market reports that Minneapolis investors Irwin Jacobs and Carl Pohlad may seek control of the company.
In September, IC announced plans to divest its Illinois Central Gulf Railroad unit by distributing shares in the division to IC holders. It estimated the railroad's market value at $250 million.
[...] Example: September 16, 2008 (cont.)<br>
slide13. Prevalence of Crash Narratives<br>
slide14. Data Financial News data:
ProQuest Wall Street Journal Eastern Edition (1984 – 2020).
231,230 articles.
Hand-collected Wall Street Journal data from 1929.
Investor survey data:
Investor Behavior Project (Shiller Survey).
19,426 responses (1989 – 2020).
Crash probability estimates, general comments, date of response, investor type (i.e. individual, institutional).
Internet search data:
Google Trends API.<br>
slide15. 1. How Do Narratives Propagate? Assess virality using internet search volumes (Da et al., 2011).
Investor interest related to crashes (CrashAttention) should be higher after seeing crash narratives in financial press.
Effects should be conditioned on general stock market attention.
Aggregate search volumes, not user-level.
Model specification:
CrashAttentiont+1 = β1 × ’87 Narrativet + β2 × StockMarketAttentiont + β3 × ’87 Narrativet × StockMarketAttentiont + Σj=15 β4,j × Xt-j+1 + Σj=15 β5,j × CrashAttentiont-j+1 + κDOWt + κMontht + ηt+1
Control variables (X): VIX, option-implied crash probabilities, market returns, market volatility, turnover.<br>
slide16. 1. Result: Propagation Evidence of positive feedback loop related to media narratives.
Other results:
Weaker effects for crash narratives from ‘29 crash (salience).
Crash narratives have positive feedback on market volatility.<br>
slide17. 2. Do Narratives Affect Investor Beliefs? Fundamental news captured by the narrative measure will likely affect individual and institutional investors similarly.
However, individuals may rely more on judgement heuristics when expertise low (Ottati & Isbell, 1996; Isbell & Wyer, 1999).
Focus on differential effect based on investor sophistication.
Model specification (by investor type):
πAdji,t = γ1 × ’87 Narrativet-1 + γ2 × StockMarketAttentiont-1 + γ3 × ’87 Narrativet-1 × StockMarketAttentiont-1 + γ4 × Wi,t-1 + λDOWt + λYear-Montht + ζi,t
Adjusted crash probabilities (πAdj): survey crash probability minus option-implied crash probability.
Control variables (W): VIX, market returns, past month market returns, market volatility, past month survey probabilities, past month stock market attention, past month crash narratives.<br>
slide18. 2. Result: Investor Crash Beliefs Individual investors adjust crash probabilities in response to crash narratives, but not institutional investors.
Other results:
Bag-of-words measures have little impact on investor assessments.
Investors “parrot” crash narratives in general assessments.<br>
slide19. 3. When Does Media Use Narratives? Pure narrativity that is plausibly exogenous to the financial markets.
Brothers Grimm Fairy Tales widely used as archetypal narrative corpora.
Stories by motif category, estimate similarity with each WSJ article.
Use the first principal component factor to capture general narrativity.
We find evidence (not shown) that journalists use narratives to make information more accessible during periods of broader readership.
Allows us to directly assess whether the crash narrative results are actually due to narrativity rather than other factors.<br>
slide20. 3. Pure Narrativity (FolkMotif) High correlation between crash and pure narratives measures: ~90%<br>
slide21. 3. Result: Narrativity, Investor Beliefs Effects isolated to component related to “pure” narratives, confirming that the main results are related to narrativity.<br>
slide22. Closing History plays an important role in press narratives by bringing attention to past “rare” events in U.S. stock market history.
Our results suggest investors rely on narrative heuristics in forming beliefs, affecting aggregate market outcomes.
Collective and individual memory processes are not mutually exclusive.
Stories may also be a mechanism to perpetuate a particularly important event in individual memory.
Interactions between collective and individual memory processes are not well understood.<br>