Causal mechanisms behind analyst anti-herding in earnings forecasts

Bernhardt, Campello and Kutsoati (2006) showed, with a nonparametric statistic that is robust to correlated signals, common earnings shocks, and systematic optimism, that sell-side analysts' earnings forecasts systematically anti-herd: forecasts above the consensus overshoot realized earnings and forecasts below it undershoot, at rates far above the unbiased benchmark in I/B/E/S data - an economically large bias that the authors conclude is strategically chosen. What produces this contrarian bias, and how much of the measured deviation could still be an artifact, remains an open attribution question. This problem asks for a causal account of the observed anti-herding: (i) an empirical decomposition that bounds the contribution of measurement artifacts - mismeasured analyst information sets (the observed consensus), classical measurement error in forecasts and consensus, firms' endogenous earnings management toward forecast positions, and analyst heterogeneity in targeted earnings concepts or posterior percentiles - and (ii) causal identification of the mechanisms generating the remaining non-artifact behavioral component, primarily strategic incentives such as relative-performance and contest incentives, compensation schemes, career and reputational concerns, and investor demand for differentiated research, discriminated against psychological biases such as the false-consensus effect. The anti-herding measurement has since been replicated across many forecaster populations and the candidate mechanisms are formally modeled, but the attribution question - how much of the overshooting is behavioral rather than artifact and which mechanism produces it - remains open in the literature.

Background

Professional equity research is produced largely by sell-side analysts, employees of brokerages who publish forecasts of the earnings per share (EPS) of the firms they cover, alongside stock recommendations. Individual forecasts and realized earnings are recorded in databases such as I/B/E/S (Institutional Brokers Estimate System), and the consensus at a given date is a summary, typically the mean, of the forecasts extant for a firm. A long literature asks whether analysts herd, meaning bias their forecasts toward the prevailing consensus, as reputational theories such as Trueman (1994) predict: an analyst who deviates from the crowd and is wrong suffers more reputational damage than one who is wrong alongside everyone else. Empirically separating herding from innocent forecast clustering is difficult because analysts share correlated information sources, economy-wide shocks move realized earnings for many firms in the same direction after forecasts are made, later forecasters legitimately observe earlier forecasts, and analysts may report different earnings concepts or percentiles of their posterior beliefs (their own belief distributions over earnings).

Bernhardt, Campello and Kutsoati (2006) - 'Who herds?' - cut through these confounds with a nonparametric statistic. Conditioning on whether a forecast $F$ lies above or below the extant consensus $\bar{F}$, define

$$\mathbf{S}=\tfrac{1}{2}\left[\Pr(F>E\mid F>\bar{F})+\Pr(F<E\mid F<\bar{F})\right],$$

where $E$ is realized earnings. If every analyst truthfully reports the same fixed percentile (such as the median) of a calibrated posterior belief about earnings, the expectation of $\mathbf{S}$ is $1/2$ regardless of correlation across analysts' signals, common unforecasted earnings shocks, information arrival, systematic optimism or pessimism, and a constant offset between observed earnings and the earnings concept analysts target; analyst-specific targeting of percentiles or earnings concepts falls outside this benchmark and is treated as an artifact channel below. Herding pushes $\mathbf{S}$ below $1/2$; strategic exaggeration away from the consensus, called anti-herding, pushes it above. On individual quarterly EPS forecasts from I/B/E/S the authors estimate $\mathbf{S}$ well above $1/2$; the magnitude of the bias, its systematic variation with analyst following, and the pattern of bias in forecast revisions lead them to conclude that the bias is economically large, strategically chosen, and not explained by analyst myopia or analyst overconfidence. The field's foundational measurement thus says analysts systematically issue contrarian, exaggerated forecasts rather than herding.

What remains open is why, and how much of the signal is real. Theory offers competing mechanisms: in Ottaviani and Sorensen (2006), winner-take-all forecasting contests make equilibrium forecasts excessively differentiated, while reputational cheap-talk biases forecasts toward the prior mean; career-concern evidence (Hong, Kubik and Solomon, 2000) shows that bold but inaccurate forecasts damage young analysts' careers, a herding force; psychological biases such as the false-consensus effect, the tendency to overestimate how similar others' information is to one's own, can also generate contrarian-looking patterns (Williams, 2013). In parallel, four identification caveats bear on any attribution of $\mathbf{S}-1/2$ to behavior: (i) the consensus an analyst actually observed is not recorded in forecast databases, which publish with lags, so the conditioning set must be reconstructed under assumptions that can themselves bias the test statistic; (ii) classical measurement error in recorded forecasts and in the consensus need not be empirically decomposed; (iii) firms manage earnings in response to forecast positions, since meeting or beating analyst expectations is rewarded (Matsumoto, 2002; Bartov, Givoly and Hayn, 2002), so realized earnings are partly an endogenous response to the very forecasts being tested; and (iv) uncontrolled heterogeneity in the targeted earnings concept (for example GAAP, Generally Accepted Accounting Principles, versus pro-forma EPS that excludes unusual items) or in the targeted posterior percentile can mimic anti-herding, even though a constant, analyst-invariant offset between the targeted concept and observed earnings is absorbed by the test's design. Distinguishing genuine behavioral contrarianism from these artifacts, and identifying which mechanism generates the non-artifact part, is the unresolved question.

Problem Statement

Using analyst-level earnings-forecast data (such as I/B/E/S), deliver a causal account of the contrarian overshooting of the consensus measured by the $\mathbf{S}$ statistic of Bernhardt, Campello and Kutsoati (2006), defined as

$$\mathbf{S}=\tfrac{1}{2}\left[\Pr(F>E\mid F>\bar{F})+\Pr(F<E\mid F<\bar{F})\right],$$

where $F$ is an analyst's forecast of a firm's earnings, $\bar{F}$ is the consensus (mean) of the forecasts extant when the forecast is issued, and $E$ is realized earnings; under truthful reporting of a common fixed percentile (such as the median) of calibrated posterior beliefs the expectation of $\mathbf{S}$ is $1/2$, while $\mathbf{S}>1/2$ indicates anti-herding. The problem has two linked components that serve one scientific objective - attributing the documented anti-herding to empirically supported mechanisms rather than to measurement artifacts:

(1) Artifact decomposition: bound, control, or empirically decompose how much of the deviation of $\mathbf{S}$ from $1/2$ is attributable to non-behavioral sources: (i) mismeasurement of the analyst's real-time information set, since the observed consensus need not be the consensus the analyst actually conditioned on; (ii) classical measurement error in recorded forecasts and in the consensus; (iii) endogenous earnings management by firms responding to analysts' forecast positions, which makes realized earnings partly a response to the forecasts under test; and (iv) heterogeneity across analysts in the targeted earnings concept (for example GAAP, Generally Accepted Accounting Principles, versus pro-forma EPS that excludes unusual items) or in the targeted posterior percentile (the point of their belief distribution over earnings that they choose to report).

(2) Mechanism identification: identify which mechanisms causally generate the remaining non-artifact behavioral component. Because Bernhardt, Campello and Kutsoati (2006) concluded that the bias is strategically chosen and ruled out analyst myopia and overconfidence, the primary candidates are strategic incentives - relative-performance or contest incentives, compensation schemes, career and reputational concerns, and investor demand for differentiated research - while psychological biases such as the false-consensus effect (overestimating how similar others' information is to one's own) are rival non-strategic explanations; the identification must separate mechanisms with observationally similar implications.

How much of the documented overshooting of $\mathbf{S}$ above $1/2$ in analyst earnings-forecast data is attributable to the non-artifact behavioral component, and which mechanisms produce it, is the open question; evidence that the deviation is instead attributable to the artifact channels would undermine the behavioral interpretation.

References

1. Bernhardt, Dan; Campello, Murillo; Kutsoati, Edward (2006). 'Who herds?' Journal of Financial Economics 80(3):657-675. DOI: 10.1016/j.jfineco.2005.07.006. https://doi.org/10.1016/j.jfineco.2005.07.006
2. Trueman, Brett (1994). 'Analyst Forecasts and Herding Behavior.' Review of Financial Studies 7(1):97-124. DOI: 10.1093/rfs/7.1.97. https://doi.org/10.1093/rfs/7.1.97
3. Hong, Harrison; Kubik, Jeffrey D.; Solomon, Amit (2000). 'Security Analysts' Career Concerns and Herding of Earnings Forecasts.' The RAND Journal of Economics 31(1):121-144. DOI: 10.2307/2601032. https://doi.org/10.2307/2601032
4. Ottaviani, Marco; Sorensen, Peter Norman (2006). 'The strategy of professional forecasting.' Journal of Financial Economics 81(2):441-466. DOI: 10.1016/j.jfineco.2005.08.002. https://doi.org/10.1016/j.jfineco.2005.08.002
5. Marinovic, Ivan; Ottaviani, Marco; Sorensen, Peter Norman (2013). 'Forecasters' Objectives and Strategies.' Handbook of Economic Forecasting, pp. 690-720. DOI: 10.1016/b978-0-444-62731-5.00012-9. https://doi.org/10.1016/b978-0-444-62731-5.00012-9
6. Williams, Jared (2013). 'Financial Analysts and the False Consensus Effect.' Journal of Accounting Research 51(4):855-907. DOI: 10.1111/1475-679x.12016. https://doi.org/10.1111/1475-679x.12016
7. Jegadeesh, Narasimhan; Kim, Woojin (2010). 'Do Analysts Herd? An Analysis of Recommendations and Market Reactions.' Review of Financial Studies 23(2):901-937. DOI: 10.1093/rfs/hhp093. https://doi.org/10.1093/rfs/hhp093
8. Naujoks, Marcel; Aretz, Kevin; Kerl, Alexander G.; Walter, Andreas (2009). 'Do German security analysts herd?' Financial Markets and Portfolio Management 23(1):3-29. DOI: 10.1007/s11408-008-0093-7. https://doi.org/10.1007/s11408-008-0093-7
9. Pierdzioch, Christian; Rülke, Jan Christoph; Stadtmann, Georg (2010). 'New evidence of anti-herding of oil-price forecasters.' Energy Economics 32(6):1456-1459. DOI: 10.1016/j.eneco.2010.05.014. https://doi.org/10.1016/j.eneco.2010.05.014
10. Pierdzioch, Christian; Rülke, Jan Christoph; Stadtmann, Georg (2012). 'House Price Forecasts, Forecaster Herding, and the Recent Crisis.' International Journal of Financial Studies 1(1):16-29. DOI: 10.3390/ijfs1010016. https://doi.org/10.3390/ijfs1010016
11. Rülke, Jan-Christoph; Silgoner, Maria; Wörz, Julia (2016). 'Herding behavior of business cycle forecasters.' International Journal of Forecasting 32(1):23-33. DOI: 10.1016/j.ijforecast.2015.02.004. https://doi.org/10.1016/j.ijforecast.2015.02.004
12. Rangvid, Jesper; Schmeling, Maik; Schrimpf, Andreas (2013). 'What do professional forecasters' stock market expectations tell us about herding, information extraction and beauty contests?' Journal of Empirical Finance 20:109-129. DOI: 10.1016/j.jempfin.2012.11.004. https://doi.org/10.1016/j.jempfin.2012.11.004
13. Blasco, Natividad; Corredor, Pilar; Ferrer, Elena (2016). 'How do analysts deal with hard-to-value firms? Evidence of herding.' Spanish Journal of Finance and Accounting 45(3):327-344. DOI: 10.1080/02102412.2016.1172196. https://doi.org/10.1080/02102412.2016.1172196
14. Tsuchiya, Yoichi (2021). 'Crises, market shocks, and herding behavior in stock price forecasts.' Empirical Economics 61(2):919-945. DOI: 10.1007/s00181-020-01894-4. https://doi.org/10.1007/s00181-020-01894-4
15. Kerl, A. G.; Pauls, T. (2014). 'Analyst herding and investor protection: a cross-country study.' Applied Financial Economics 24(8):533-542. DOI: 10.1080/09603107.2014.889800. https://doi.org/10.1080/09603107.2014.889800
16. Spyrou, Spyros (2013). 'Herding in financial markets: a review of the literature.' Review of Behavioral Finance 5(2):175-194. DOI: 10.1108/rbf-02-2013-0009. https://doi.org/10.1108/rbf-02-2013-0009
17. Matsumoto, Dawn A. (2002). 'Management's Incentives to Avoid Negative Earnings Surprises.' The Accounting Review 77(3):483-514. DOI: 10.2308/accr.2002.77.3.483. https://doi.org/10.2308/accr.2002.77.3.483
18. Bartov, Eli; Givoly, Dan; Hayn, Carla (2002). 'The rewards to meeting or beating earnings expectations.' Journal of Accounting and Economics 33(2):173-204. DOI: 10.1016/s0165-4101(02)00045-9. https://doi.org/10.1016/s0165-4101(02)00045-9
19. Ghimire, Binam; Reveley, Callum; Shanaev, Savva; Panta, Humnath (2023). 'Analyst herding - whether, why, and when? Two new tests for herding detection in target forecast prices.' Economics and Business Review 9(4). DOI: 10.18559/ebr.2023.4.892. https://doi.org/10.18559/ebr.2023.4.892
20. Nguyen, Lan Thi Mai; Song, Jun Myung; Truong, Cameron (2026). 'Bilateral political relations and analyst herding.' The British Accounting Review, article 101837. DOI: 10.1016/j.bar.2026.101837. https://doi.org/10.1016/j.bar.2026.101837