Week 7: Herding

1. Herding behavior: introduction

This lecture is about how investors are influenced by the actions of others and end up acting collectively. The central question: if everyone has information, why do markets so often move "like a herd"? Three examples from the slides:

• the 2021 "GameStop rally", coordinated via social media,
• the mass withdrawal from funds in 2008,
• and the dot-com bubble of 1999 to 2000.

Herding is the social mechanism behind many bubbles and crises. Last week we saw that sentiment is a collective mood; herding is the way that mood is turned into trades that all point in the same direction.

Why does this happen?

The slides give three reasons why investors end up trading in the same direction:

• They are irrational: the GameStop "YOLO" (you only live once) traders bought because others did and because it was exciting.
• They react similarly to common information about fundamentals: when tariffs are announced, many investors sell exporters at the same time. This looks like herding but is a shared reaction to news.
• They infer things from what others did in the prior period: think of retail investors who mirror the published trades of U.S. politicians like Nancy Pelosi on the assumption that she has inside knowledge.

From this, the literature distinguishes three types of herding:

1. Informational: you copy others because their actions reveal information.
2. Reputational or career-based: you copy others to protect your job or reputation.
3. Emotional or social (behavioral): you copy others because of feelings such as FOMO or the comfort of the crowd.

2. The five foundational papers

Five studies form the "core infrastructure" of the herding literature:

Scharfstein and Stein (1990): herding from career concerns

Fund managers copy one another to safeguard their reputation. Being "wrong with everyone" is preferred to being "right alone". This follows from the way managers are judged. A loss on a stock that all the other funds owned too gets read by the employer as bad luck that affected everybody. A loss on a contrarian position nobody else took reads as poor judgment and can cost her the job. The safe career move is therefore to stick to the benchmark and to what peers do: "if everyone is buying, it's safe for me to buy". The result is herding that is "rational" for the manager, with no psychological bias involved. The slides label it a top-tier paper; it appeared in the American Economic Review (1990).

Banerjee (1992): sequential herding

Investors act one after another and observe the decisions of those before them. The mechanism:

1. Each investor has private information (a signal).
2. Each observes the decisions of previous investors.
3. If two or three earlier investors made the same choice, the next investor ignores her own signal and follows them.

A classic illustration is choosing between two restaurants. You read a review favouring restaurant A, but you see that B is full and A is empty. You conclude the people in B probably know something and go to B. The diners in B may all have done the same reasoning, so the crowd can be wrong even though everyone acted sensibly.

Bikhchandani, Hirshleifer and Welch (1992): information cascades

This is the classic model of information cascades and the most famous paper on herding (more than 10,000 citations). Its core:

• Public information consists of the actions of others.
• As more people act, private information becomes less important.
• A cascade arises: everyone imitates, without any new fundamental information.

The slide's chart shows the weight of public information rising with each investor in the sequence: about 0.2 for investor 1, 0.4 for investor 2, 0.7 for investor 3, 0.9 for investor 4 and 1.0 for investor 5. By then private information has no weight at all.

Lakonishok, Shleifer and Vishny (1992): institutional herding

LSV introduced the first empirical measure of institutional herding, applied to the quarterly trades of U.S. pension and mutual fund managers:

H_j,t = |p_j,t − E(p_j,t)| − AF_j,t

Results: herding is moderate (0.03 to 0.06) but persistent over time, more pronounced in small caps and "risky" sectors, and it leads to short-term mispricing. The paper laid the foundation for the literature on institutional herding. (Section 7 works through the formula in detail.)

Chang, Cheng and Khorana (2000): CSAD market herding

CCK introduced the cross-sectional absolute deviation (CSAD), the most widely used herding measure today:

CSAD_t = (1/N) Σ |R_i,t − R_m,t|

CSAD_t = α + γ₁|R_m,t| + γ₂R_m,t² + ε_t

• R_i,t = return of stock i on day t; R_m,t = market return (the average across stocks) on day t; N = number of stocks.
• CSAD measures how far individual stock returns are spread out around the market return on a given day. Low CSAD means stocks move together.
• α, γ₁ and γ₂ are coefficients estimated with a regression over many days.

Intuition. Rational asset pricing models (like the CAPM) imply that stocks have different sensitivities (betas) to the market. On a day when the market moves a lot, stocks with high and low betas move by very different amounts, so dispersion should rise linearly with the size of the market move. With herding, investors suppress their own views and follow the market, so stocks move together and dispersion rises less than linearly, or even falls. This shows up as a negative γ₂: if γ₂ < 0, dispersion does not increase as much as market moves get larger, so investors move more "together" than fundamentals would predict.

The two charts on the slides show this shape. In the first, CSAD rises from 0.5 at |R_m| = 0 to 0.9 at 1 and 1.1 at 2, but then flattens at about 1.15 to 1.17 for |R_m| of 3 and 4. In the second, dispersion rises from 0.5 to about 3.2 as |R_m| goes from 0 to 5, but along a curve that bends down. In a normal market dispersion increases linearly; with herding the line "flattens" (concavity).

How the five papers fit together

• Theoretical: Banerjee (1992) and Bikhchandani et al. (1992) explain informational herding; Scharfstein and Stein (1990) explain reputational herding.
• Empirical: LSV (1992) measures institutional herding in funds; CCK (2000) measures market-wide herding in returns.

Conclusion: herding is multidimensional:

Herding = Information (Banerjee / BHW) + Reputation (Scharfstein & Stein) + Behavioral (psychology)

Modern research uses all five as an integrated framework: theory tells you why people would herd, the measures tell you whether they do.

3. Additional studies

1. Graham (1999) presents a model in which analysts (in his paper, investment newsletter writers) are much more likely to herd when they have a high reputation, to protect that reputation, or low ability, to hide their low ability. Analysts in the middle have the most reason to deviate.
2. Cipriani and Guarino (2005) ran a laboratory experiment that simulated a capital market. When individuals trade based on information in a market without distortions (prices adjust to the order flow), herding occurs very rarely. The reason: in a market, the price already reflects the information in previous trades, so following the crowd at a higher price is no longer attractive. This suggests that the cascades of the theory need frictions to appear in markets.
3. Brunnermeier and Nagel (2004) found that hedge funds during the NASDAQ (dot-com) bubble did not fight the bubble. They rode it: they invested heavily in technology stocks, exploiting the sentiment of retail investors, but exited their positions significantly earlier than other investors, before the crash. Smart money herded on purpose, for profit.

4. Bubbles and anti-bubbles

1. It is reasonable to associate bubbles with herding behavior.
2. During a bubble, prices rise sharply and lose their relationship with fundamentals.
3. Investing in specific "popular" stocks during a bubble means that investors make similar choices and follow one another.
4. When the bubble bursts and investors exit en masse, herding exists in the opposite direction.
5. Johansen and Sornette (1999) call this an anti-bubble: a downward spiral driven by the same imitative behavior as the bubble. Their research on the Japanese Nikkei index after its 1989 peak found such patterns.

5. Spurious versus intentional herding

Bikhchandani and Sharma (2000) make a distinction that is central for empirical work:

• Spurious herding: investors make similar decisions because they face common information. Example: the ECB unexpectedly raises rates and all investors sell rate-sensitive stocks.
• Intentional herding: investors deliberately follow other investors, ignoring their own information.

Spurious herding is consistent with market efficiency: prices move because the fundamentals changed. Intentional herding may create systemic risk, because prices can move away from value and because cascades are fragile and can reverse suddenly.

6. Psychological basis of herding

• Social proof: people believe that "if everyone is doing something, it must be right".
• Conformity bias: adjusting to the majority's views for acceptance and safety (as in Asch's famous experiments, where people gave obviously wrong answers to match the group).
• Fear of missing out (FOMO): fear of being left out of a rising market while friends get rich.
• Loss aversion: fear of losses makes investors copy others during crises; losing together hurts less than losing alone.

The herd works as a psychological survival mechanism in markets. In evolutionary terms, staying with the group was usually safer than standing alone.

Informational herding revisited: Banerjee and BHW

In Banerjee's model investors act sequentially; when enough predecessors buy, the next investor ignores her own information and follows ("if everyone is buying, why should I sell?"), and an information cascade is created. The logic of the first few investors becomes a collective illusion of correctness. In the BHW model, investors have private information but also observe others; once public information (others' actions) outweighs private information, a cascade arises and the market moves in one direction without new fundamental information. The key conclusion: even fully rational investors can herd. Herding does not require stupidity, only the rational use of other people's actions as information.

Reputational herding revisited: Scharfstein and Stein

Fund managers prefer to be wrong together than wrong alone; if everyone is buying an asset, it is "safe" for your reputation to do the same; career incentives create uniform investment decisions. Herding is therefore not necessarily irrational from an individual standpoint: it is strategically rational, even if the outcome for the market (and for the fund's clients) is bad.

Distinguishing herding from common information

• When everyone buys because there is common fundamental news, that is not herding.
• When they buy without such news, simply because "everyone is buying", that is herding.

In summary: correlation in behavior ≠ correlation in information. The main challenge for empirical studies is to separate the two. This is why the CSAD approach compares dispersion to what a rational pricing model predicts, and why some studies (like Galariotis et al., 2015, below) split herding into a part driven by fundamentals and a part that is not.

7. Herding in real time: GameStop

In 2021, on Reddit's WallStreetBets forum, millions of investors bought GameStop simultaneously, not because of fundamentals but because of a collective belief: "we will defeat the hedge funds". Similar episodes followed in crypto (Dogecoin, Shiba Inu). The herd now forms online: social media are the new "market square".

What happened (December 2020 to February 2021)

• GameStop was a declining company (physical video game stores).
• Hedge funds had massive short positions: more than 120% of the available float was sold short (possible because borrowed shares can be lent out again).
• The idea caught on at WallStreetBets: if everybody bought at once, the group could squeeze the hedge funds out of their short positions.
• Thousands of retail investors started buying at the same time.
• Robinhood halted buy orders in GameStop since it was unable to execute them (the lecturer adds a wry "cough, Citadel, cough", a reference to the market maker processing Robinhood's order flow).
• Result: the price of GME rose from about $17 to over $480 (pre-split prices) within three weeks.

The "From Nada to Prada" slide shows GME at $212.94 in June 2021, up about 4,272% over one year, after a first spike near $350 (daily close) in late January 2021.

The mechanism: herding plus short squeeze

1. Herding behavior: Reddit, Twitter and TikTok created a narrative ("YOLO", "diamond hands") and retail investors imitated each other without fundamentals.
2. Short squeeze: to close their short positions the hedge funds were forced to buy shares. Heavy retail buying combined with short covering made the price explode.

Why GameStop is a behavioral phenomenon

• FOMO: the fear of missing out on the rally.
• Attention-driven buying: massive media coverage pushed up trading volume.
• Herding: imitation on a mass scale through social media.
• Overconfidence: an exaggerated belief that "it won't fall".
• Narrative contagion (Shiller): the story of small investors beating the hedge funds.
• Disposition effect: investors would not sell ("diamond hands"). This is an unusual form: the disposition effect normally means selling winners too soon, whereas here holding on was a social commitment.

Result: as soon as the narrative fell apart the price crashed, and those who came in late suffered big losses. GameStop illustrates that the herd can both build and wipe out value.

8. Emotional herding and the herding cycle

In emotional herding, decisions are based on emotion and atmosphere, not on information. The "market mood" spreads through media and investment networks, and when sentiment changes, the herd reverses direction. The link with last week: herding is the behavioral manifestation of investor sentiment.

The herding cycle has four stages: initial trend, imitation, trend reinforcement and collapse or panic, after which the cycle restarts. The herd amplifies market momentum until collapse.

The slide titled "'Diversification' and other hilarious jokes to tell yourself" shows six major stock indices (S&P 500, FTSE 100, DAX, CAC 40, Nikkei 225 and India's Nifty 50) from November 2024 to November 2025. They all fell sharply together in April 2025 and recovered together. The point: when global investors herd, markets that should diversify each other become highly correlated, exactly when diversification is needed most.

9. Measuring herding empirically

The problem: how do we detect herding in return or portfolio data? The main methods:

1. Cross-sectional dispersion models: Christie and Huang (1995), Chang, Cheng and Khorana (2000). They use stock returns.
2. Holdings-based measures: Lakonishok, Shleifer and Vishny (1992) for institutional investors. They use portfolio holdings and trades.
3. Regression and dummy-based approaches: these test for symmetric co-movement toward the market.

The common idea: if stock returns (or trades) move together more than fundamentals can explain, there is herding.

Christie and Huang (1995): cross-sectional standard deviation (CSSD)

CSSD_t = √[ (1/(N − 1)) Σ_i=1^N (R_i,t − R_m,t)² ]

R_i,t is the return of stock i and R_m,t the market return. CSSD is the standard deviation of stock returns around the market on one day. If herding exists, dispersion should decrease during extreme market movements, because that is when investors are most tempted to follow the crowd.

To test this they estimate:

CSSD_t = α + β₁D_L,t + β₂D_U,t + ε_t

D_L,t is a dummy (a 0/1 variable) equal to 1 on days with an extreme negative market return (for example the lowest 1% or 5% of days), and D_U,t equals 1 on extreme positive days. α is the average dispersion on normal days. Rational pricing predicts higher dispersion on extreme days (positive βs). A negative β is evidence of herding.

The CSAD approach of CCK (section 2) improved on CSSD in two ways: absolute deviations are less sensitive to outliers than squared ones, and the regression on |R_m| and R_m² uses all days instead of requiring an arbitrary cut-off for "extreme" days.

The LSV measure in detail

H_j,t = |p_j,t − E(p_j,t)| − AF_j,t

• p_j,t = share of institutions that are buying stock j in quarter t = buyers / (buyers + sellers).
• E(p_j,t) = expected proportion if there were no coordinated behavior; in practice the average proportion of buyers across all stocks in that quarter.
• AF_j,t = adjustment factor for random variation: even when institutions trade independently, |p − E(p)| will not be exactly zero by chance, especially with few traders. AF is the expected value of |p − E(p)| under independent trading.

Intuition: the more institutions move in the same direction (far more buyers than sellers, or the reverse), the higher H.

Findings from LSV and CSAD

• Lakonishok et al. (1992): U.S. institutions herd to a small or moderate degree, more so in small stocks and stocks with little coverage (only a few analysts follow them).
• Chang et al. (2000): significant herding in emerging markets (South Korea, Taiwan) and partly in Hong Kong, weaker or absent in developed markets (U.S., Japan).

Herding is stronger when information is limited and uncertainty is high. When you know little, others' actions are more informative to you.

Sias (2004)

• Idea: when institutions herd, the share of institutions buying a stock this quarter should move together with the share that bought it one quarter earlier, meaning they follow one another into the same names over time.
• For every institution and every stock, he records the holding as a share of shares outstanding at both the start and the end of the quarter.
• If an institution increases (decreases) its holding, it is classified as a buyer (seller). For each stock and quarter, the percentage of buyers is computed: the "raw fraction of institutions buying".
• He then measures the cross-sectional correlation between this quarter's fraction and last quarter's fraction. A positive correlation indicates herding.
• Holmes, Kallinterakis and Ferreira (2013) applied this approach to the Portuguese market and found institutional herding.

Institutional versus retail herding

• Institutional investors: herding that is rational or career-driven (fear of trailing the benchmark), rooted in reputational incentives (Scharfstein and Stein, 1990).
• Retail investors: emotional or social herding (FOMO, sentiment contagion), reacting to social media and news.

Whether herding is rational or psychological depends on which kind of investor is involved.

10. Consequences of herding

Herding and returns

• Herding creates return correlation and reduces the benefits of diversification.
• It increases volatility and reduces market efficiency.
• It creates momentum in the short term (the herd keeps buying winners) and reversal in the long term (prices return to value).
• When the herd reverses, we see mass corrections or "fire sales" (forced selling at very low prices).

Global evidence

• Emerging markets: strong herding (low transparency, limited information).
• Fund flows: institutional investors imitate one another.
• FX, bonds and crypto: coordinated movements during panic episodes.

The slide's stylised CSAD chart moves between about 0.2 and 0.5 over seven periods, with a fall at period 4 and a sharp fall from 0.5 to 0.2 at period 7. The message: a drop in CSAD (stocks suddenly moving together) indicates strong herding during crises.

Herding and bubbles

• The more people take part, the stronger the upward move gets.
• Successes of "early adopters" serve as examples and give others social validation.
• Prices lose their link with fundamentals: the bubble becomes self-fulfilling.
• Once the herd turns, the bubble bursts suddenly.

Shiller: "Bubbles are social episodes of collective illusion."

11. Studying herding in the stock market: Galariotis, Rong and Spyrou (2015)

The standard approach for a study (and a good template for a student project):

• apply the CSAD model to stock data (for example the Athens Stock Exchange or the S&P 500),
• examine crisis periods separately (the global financial crisis, COVID, the energy crisis),
• if γ₂ < 0 in CSAD_t = α + γ₁|R_m,t| + γ₂R_m,t², herding is present.

Galariotis, Rong and Spyrou (2015), "Herding on fundamental information", work with daily U.S. and U.K. stock return data covering 1989 to 2011. They decompose CSAD into a component that fundamental risk factors explain and a residual component, which lets them test herding driven by fundamental information (spurious) apart from herding that has nothing to do with fundamentals.

Main findings for the U.S.:

• On "normal" trading days there is no systematic or permanent herding: return dispersion did not shrink as if everyone moved together.
• Strong evidence of herding appears on days with important macroeconomic information: FOMC (Fed interest rate) announcements, employment releases, and volatility or crisis announcements.
• Herding in the U.S. is therefore conditional: activated under high common information or stress, not by general imitation.
• It is more pronounced in small caps, in high-risk / low-information stocks and in times of high uncertainty.

Their CSAD chart for the U.S. (1989 to 2011) shows daily CSAD mostly between about 0.01 and 0.02, with large spikes up to about 0.06 around 2000 to 2002 (dot-com crash) and 2008 to 2009 (financial crisis). Note that crises show high CSAD in absolute terms because markets move a lot; herding is about whether CSAD is lower than the size of the market move would predict.

Interpretation and lessons

• In mature markets like the U.S., investors probably have better information and stronger arbitrage, so herding is less persistent than in emerging markets.
• When very important common information appears (earnings announcements, monetary policy), investors may "synchronise".
• For investors and managers: take extra care on announcement days, since everybody gets the same news then and may simply go along with the crowd.
• For regulators: crowd reactions do not happen all the time, but when they surface on critical days they may point to systemic risk.

Conclusion: herding is not the rule in the U.S.; it switches on only under particular conditions.

Conclusions so far

• Herding ranks among the most stable behavioral patterns found in markets.
• Empirical measures such as CSAD and LSV show it is present across countries and over time.
• It produces volatility, correlations and pricing anomalies.
• Keeping track of herding is key both for risk management and for understanding crises.

12. Behavioral interpretations and practice

• Social proof: considering correct what the majority does.
• Fear of isolation: being afraid of standing alone as "the only one doing the opposite".
• Information cascade: the decisions of the first movers disproportionately affect later ones.
• Loss aversion: losing alongside the crowd feels better than losing on our own.
• Overconfidence: the illusion that "I will exit before the others".

The lecture concludes that herding is not just irrational: it is socially adaptive. Real-world cases: the dot-com bubble (institutions and retail alike piled into tech without any fundamental basis), the fall of Lehman Brothers in 2008 (money fled every type of fund in a "flight to safety" toward government bonds and cash) and the 2021 crypto rally (herding fuelled by social media, memes and FOMO). The herd converts information into a flowing narrative.

Herding and investment strategies

• Herd-following: join the trend (momentum trading).
• Contrarian: invest against the herd (reversal strategies). Contrarians "buy fear" and "sell euphoria".

The slide's stylised chart shows contrarian returns falling as sentiment rises: about 8 at sentiment −2, about 2 at 0 and about −2 at +2. The profitability of a contrarian strategy increases when the herd is more extreme, because extreme herding means larger mispricing to be corrected. This links to De Bondt and Thaler's (1985) loser-minus-winner evidence from earlier weeks. The two strategies are not contradictory: momentum profits from the short-term continuation that herding creates; contrarian profits from the long-term reversal.

Herding and risk management

• Herding increases systemic risk: everyone holds the same assets.
• It increases correlation between securities and reduces the effectiveness of diversification.
• When all investors try to sell simultaneously, it gives rise to liquidity spirals.

Mitigation: monitor correlations, run stress tests on crowded trades (positions held by many funds), and monitor institutional flows and sentiment indicators.

Herding and artificial intelligence

Social media produce "digital" herding in real time. Machine learning can pick up these patterns in trading or tweet data; predictive models like CNNs (convolutional neural networks) and LSTMs (long short-term memory networks) are able to spot synchronised investor behavior. Behavioral finance is moving into an age of data-driven herding.

How we detect herding in data

• Return or fund-flow correlations jump sharply.
• Signals in news or social media (sentiment or wording that looks alike).
• Hedge funds piling into "crowded trades" (all holding the same assets).
• When the market moves strongly, cross-sectional dispersion falls (a low CSAD).

Linked to sentiment analysis: Herding = f(correlated behavior, high sentiment). Herding is most probable when correlated behavior and extreme sentiment occur together.

Political and regulatory dimension

Regulators care about herding because of systemic risk. They supervise crowded trades, run stress tests for funds, monitor capital flows and the behavior of ETFs and passive investors (passive funds all buy the same index stocks, a structural form of herding), and use interventions such as circuit breakers (temporary trading halts after large price falls) during mass panic.

13. Analysts and the "Cramer effect"

Herding among analysts

• Cote and Sanders (1997) show that the need to protect an analyst's reputation leads to herding.
• Hirshleifer, Subrahmanyam and Titman (1994) find herding in analyst announcements.
• Hong, Kubik and Solomon (2000) find that inexperienced analysts whose forecasts deviate from the consensus are more likely to lose their jobs than experienced analysts whose forecasts lie at the mean. Young analysts therefore have a strong incentive to stay close to the consensus, which is exactly the Scharfstein and Stein mechanism.

Jim Cramer and Inverse Cramer

Jim Cramer presents "Mad Money" on CNBC and is TV's best-known (and noisiest) stock picker, handing out "buy", "sell" or "hold" calls every day. The lecture sees his influence as herding plus a celebrity effect: retail investors respond more to personalities than to analysis, and quite a few of his picks went badly wrong. A CNBC screenshot on one slide reads: "Bitcoin is up 84% since Jim Cramer said 'I would sell my Bitcoin' 9 months ago". On X, the account @CramerTracker keeps track of his calls.

The "Inverse Cramer" strategy: if his picks cause short-run overvaluation (viewers herd into the stock) followed by long-run underperformance, betting against him could pay off. Alternatively, the lecturer remarks, he might just be a poor analyst. In June 2022 a headline claimed that "The Inverse Cramer ETF is still beating the S&P 500 by nearly 7 percent". Yet the Inverse Cramer Tracker ETF (ticker SJIM), started in 2023, performed badly: indexed at 100 in late February 2023, it had dropped to about 90 by November 2023, while the S&P 500 climbed to about 114. In the end the ETF was shut down. The takeaway for researchers: an anecdote or a brief backtest does not prove a durable anomaly; a strategy has to hold up out of sample.

14. Conclusion

• Every institutional investor or risk manager has to understand herding.
• It explains why markets "move together" more than fundamentals justify.
• Its roots are social, psychological and professional incentives.
• The result can be bubbles and crashes, but also investment opportunities.

The lecture ends with an honest look at the evidence:

• For institutional investors, some studies find limited evidence of herding while others find significant evidence.
• Many studies find significant herding by analysts for various reasons, but later studies conclude that analysts sometimes try to differentiate themselves from the herd (anti-herding, for example to get noticed).
• The methods have limitations. The LSV statistic may show herding even when only a small number of institutions are active in a stock: with 3 traders, all buying by chance is quite likely.
• LSV also cannot distinguish managers who trade because they have (the same) information from managers who imitate others: it cannot separate spurious from intentional herding.