Earnings Surprise and Analyst Revision Overlay: Joint Signal Predictive Power

Overview

Earnings surprises and analyst revisions represent two distinct information channels into equity returns: one backward-looking (actual earnings vs. consensus forecast), the other forward-looking (shifts in professional opinion). The empirical evidence demonstrates that these signals are neither fully redundant nor simply additive; instead, their joint predictive power depends critically on information uncertainty, analyst forecast bias, and the market's asymmetric processing of earnings news. Analyst revision changes emerge as a robust, orthogonal predictor, while earnings surprises drive initial underreaction dynamics that revisions can either amplify or correct.

Key Findings

1. Analyst Revisions Contain Orthogonal Information to Earnings Surprises

[P1] provides the foundational empirical result: the quarterly change in consensus recommendations is a robust return predictor that "appears to contain information orthogonal to a large range of other predictive variables." This is critical—revisions are not merely echoing the earnings surprise signal. The study shows that while the level of consensus recommendations adds value only among stocks with favorable quantitative characteristics (value stocks, positive momentum), the change in recommendations is a general return predictor. This suggests that analyst revisions capture forward-looking sentiment shifts that are independent of the contemporaneous earnings surprise.

Quantitative implication: The orthogonality of revision changes to earnings surprises means a joint model should include both as separate regressors without multicollinearity concerns, and their coefficients should remain statistically significant when both are included.

2. Information Uncertainty Modulates the Earnings Surprise Effect

[P6] demonstrates that the post-earnings-announcement drift (PEAD) anomaly—the tendency for stock prices to drift in the direction of the earnings surprise for weeks after announcement—is substantially mediated by information uncertainty (IU). The key findings:

  • Unexpected earnings (UE) signals with greater information uncertainty experience more muted initial market reactions
  • Extreme UE portfolios contain securities with higher IU than non-extreme portfolios
  • Within extreme UE portfolios, high-IU securities earn larger abnormal returns than low-IU securities

This reveals a critical interaction: when information uncertainty is high, the market underreacts more severely to earnings surprises, creating larger drift opportunities. Conversely, when IU is low (high analyst coverage, transparent business models), the initial reaction is sharper and drift is smaller. Analyst revisions, which typically follow earnings announcements, may be more impactful in high-IU environments where the initial market reaction was muted.

Quantitative implication: The PEAD effect is conditional on IU. A joint model should include an interaction term between earnings surprise magnitude and information uncertainty, and analyst revisions may have stronger predictive power in high-IU subsamples.

3. Analyst Forecast Pessimism Creates Predictable Earnings Surprise Patterns

[P9] reveals a critical asymmetry: investors do not fully unravel predictable pessimism in sell-side analysts' earnings forecasts. The study shows that:

  • Measures of prior consensus forecast pessimism are predictive of both the sign of earnings surprises and stock returns around announcements
  • Firms with high probability of forecast pessimism experience significantly higher announcement returns than those with low probability
  • This mispricing is driven by pessimism in short-term forecasts, not optimism in longer-term forecasts

This finding has profound implications for the joint signal: if analysts are systematically pessimistic in their near-term forecasts, then positive earnings surprises are more likely to occur, and analyst revisions upward will be more common. The earnings surprise and the subsequent analyst revision are not independent—they are linked through the bias in the initial forecast. A positive earnings surprise is more likely to trigger upward revisions, creating a reinforcing signal.

Quantitative implication: The joint effect of earnings surprise + upward revision may be multiplicative rather than additive, because the revision is partially predictable from the surprise itself when forecast pessimism is present. However, the magnitude of the revision (how much analysts revise) may contain additional information beyond the sign of the surprise.

4. Earnings Surprise Magnitude and Market Underreaction

[P2] provides the theoretical foundation for understanding why earnings surprises alone are insufficient predictors. The theory of investor overconfidence and biased self-attribution predicts:

  • Negative long-lag autocorrelations (overreaction to past information)
  • Positive short-lag autocorrelations ("momentum")
  • Short-run earnings drift

The implication is that earnings surprises trigger underreaction in the short run (days to weeks), followed by drift. Analyst revisions, which typically occur within days of the earnings announcement, can either reinforce the underreaction (if revisions are also muted) or accelerate price discovery (if revisions are sharp). The joint signal's power depends on whether revisions amplify or dampen the initial earnings surprise reaction.

5. Text-Based Earnings Signals Outperform Numerical Surprises Alone

[P8] introduces a critical methodological finding: a text-based measure of earnings announcement surprises (SUE.txt, constructed from earnings call transcripts) generates post-earnings-announcement drift larger than the classic PEAD based on numerical earnings surprises alone. Importantly, "the magnitude of PEAD.txt is considerable even in recent years when the classic PEAD is close to 0."

This suggests that the content of earnings announcements—the narrative and discussion around the numbers—contains predictive information that numerical surprises miss. Analyst revisions, which are typically informed by both the numerical surprise and the qualitative discussion, may capture this additional signal. A joint model incorporating both numerical earnings surprises and text-based sentiment (or analyst revisions, which implicitly incorporate text analysis) would outperform either alone.

Quantitative implication: The predictive power of earnings surprises has declined over time (as markets have become more efficient at processing numerical surprises), but text-based signals remain robust. Analyst revisions, which incorporate qualitative information, may be the mechanism through which this text-based signal is transmitted to prices.

6. Analyst Report Readability and Information Precision

[P3] demonstrates that analyst report readability is associated with stronger trading volume reactions. Reports from higher-ability analysts are more readable, and "trading volume reactions increase with the readability of analysts' text, consistent with theoretical models that predict that more precise information (and hence more informative signals) results in investors' initiating trades."

This finding suggests that the quality of analyst revisions matters. A revision from a high-ability analyst (signaled by report readability) will trigger stronger market reactions than a revision from a low-ability analyst. In a joint model, the predictive power of analyst revisions should be weighted by analyst ability or report quality.

7. Earnings Announcement Language and Forward Returns

[P5] shows that optimistic and pessimistic language in earnings press releases predicts future firm performance after controlling for the earnings surprise. The study finds "a positive (negative) association between optimistic (pessimistic) language usage and future firm performance and a significant incremental market response to optimistic and pessimistic language."

This is a direct test of whether information beyond the numerical surprise adds predictive power. The answer is yes—the tone and language of the earnings announcement contain information orthogonal to the surprise itself. Analyst revisions, which are informed by this language, should capture this signal. However, the market's initial response to the language may be incomplete, creating drift opportunities.

Quantitative implication: A joint model should include a language sentiment measure alongside earnings surprise and analyst revision. The three signals may be partially redundant, but each contains unique information.

8. Insider Trading and Earnings Surprise Interpretation

[P10] provides evidence that insider trading after earnings announcements influences the post-earnings-announcement drift. Contrarian insider trades mitigate underreaction, while confirmatory trades allow drift to continue. This suggests that the market's interpretation of an earnings surprise—whether it represents a transitory or permanent change—is not immediately clear and is resolved through subsequent trading activity.

Analyst revisions, which typically occur within days of the announcement, may serve a similar function: they provide professional interpretation of whether the earnings surprise represents a fundamental shift in the firm's prospects. A revision upward signals that analysts believe the surprise is permanent; a revision downward signals skepticism about sustainability.

Limitations and Caveats

1. Sample Period and Market Regime Sensitivity

The papers span 1998–2009 (most data), with [P8] extending to recent years. The PEAD anomaly has weakened substantially in recent decades, particularly for numerical earnings surprises. The joint signal's predictive power may have deteriorated similarly, especially if analyst revisions have become more efficient at incorporating earnings surprises. The evidence on text-based signals ([P8]) suggests that the frontier of predictability has shifted from numerical surprises to qualitative content, but the joint effect of numerical surprise + analyst revision in recent data is not directly quantified.

2. Analyst Bias and Forecast Pessimism Are Time-Varying

[P9] documents systematic pessimism in short-term analyst forecasts, but this bias is not constant across market conditions, firm characteristics, or time periods. The predictability of earnings surprises from forecast pessimism may be weaker in periods when analysts are less biased (e.g., during market downturns when pessimism is already priced in). The joint signal's power depends on the current level of analyst bias, which is not directly observable.

3. Information Uncertainty Is Endogenous

[P6] shows that IU modulates the earnings surprise effect, but IU itself may be correlated with analyst revision behavior. Firms with high IU may have fewer analysts, less frequent revisions, and larger revisions when they do occur. The interaction between earnings surprise, analyst revision, and information uncertainty is complex and may not be captured by simple additive or multiplicative models.

4. Recommendation Level vs. Revision Change

[P1] distinguishes between the level of consensus recommendations (which adds value only among favorable stocks) and the change in recommendations (which is a robust predictor). This distinction is crucial but often conflated in practice. The joint signal's power depends on whether one uses recommendation levels or changes, and the evidence suggests changes are more robust. However, the interaction between earnings surprise and recommendation change (as opposed to level) is not directly tested.

5. Causality and Reverse Causality

The papers document correlations between analyst revisions and returns, but causality is not definitively established. Do analyst revisions cause returns to move, or do analysts revise in response to price movements? [P1] argues that revisions contain orthogonal information, but reverse causality cannot be ruled out. A joint model should account for the possibility that analyst revisions are partially endogenous to price movements.

6. Multicollinearity in Joint Models

While [P1] argues that revision changes are orthogonal to earnings surprises, this orthogonality may not hold in all subsamples or time periods. In high-IU environments, analysts may revise more sharply in response to earnings surprises, creating multicollinearity. The joint model's stability across subsamples is not guaranteed.

Practical Implications for Quant Practitioners

1. Build Separate Signals, Then Combine

The evidence strongly supports treating earnings surprises and analyst revisions as distinct signals. A machine learning model should include both as separate features, not collapse them into a single "earnings news" variable. [P1] demonstrates that revision changes are orthogonal to earnings surprises, so the joint model should have lower multicollinearity than expected.

Implementation: Use standardized unexpected earnings (SUE) as one feature and the change in consensus recommendation (or the magnitude of analyst revisions) as a separate feature. Include interaction terms between the two, as the joint effect may be multiplicative in high-IU environments.

2. Weight Analyst Revisions by Ability and Readability

Not all analyst revisions are equally informative. [P3] shows that revisions from high-ability analysts (signaled by report readability) trigger stronger market reactions. A practical model should weight revisions by analyst ability, coverage breadth, or historical accuracy.

Implementation: Construct a weighted revision signal where revisions from top-quartile analysts (by historical accuracy or report readability) receive higher weights. This may improve out-of-sample predictive power.

3. Condition on Information Uncertainty

[P6] demonstrates that the earnings surprise effect is substantially larger in high-IU environments. A practical model should include information uncertainty as a moderating variable. High-IU stocks with large earnings surprises and upward revisions may offer the strongest signal.

Implementation: Estimate information uncertainty using idiosyncratic volatility, analyst coverage, or bid-ask spreads. Segment the portfolio into high-IU and low-IU subsamples and apply different thresholds for earnings surprise and revision signals in each.

4. Exploit Analyst Forecast Pessimism

[P9] shows that analyst forecast pessimism is predictive of earnings surprises and announcement returns. A practical model can exploit this by identifying firms with high forecast pessimism (e.g., consensus forecast well below historical earnings) and expecting larger positive surprises and upward revisions.

Implementation: Calculate the ratio of consensus forecast to historical earnings (or analyst forecast dispersion as a proxy for pessimism). Firms with low ratios are more likely to surprise positively. Combine this with actual earnings surprises and revisions for a three-signal model.

5. Incorporate Text-Based Signals

[P8] demonstrates that text-based earnings signals (SUE.txt) outperform numerical surprises, especially in recent years. Analyst revisions implicitly incorporate text analysis, but a practical model could explicitly include sentiment measures from earnings calls or analyst reports.

Implementation: Use natural language processing to extract sentiment from earnings call transcripts or analyst reports. Combine this with numerical earnings surprises and analyst revisions for a richer signal. The text-based signal may be particularly valuable in low-IU environments where numerical surprises are quickly priced in.

6. Exploit the PEAD Anomaly with Joint Signals

[P6] shows that PEAD is larger for high-IU securities with extreme earnings surprises. A practical model can exploit this by constructing a post-announcement drift strategy that combines earnings surprises, analyst revisions, and information uncertainty.

Implementation: On the day of earnings announcement, identify stocks with large earnings surprises (top/bottom decile), high information uncertainty, and upward/downward analyst revisions. Hold these positions for 20–60 trading days to capture drift. The joint signal (surprise + revision + IU) should outperform any single signal alone.

7. Account for Time-Varying Analyst Bias

[P9] documents systematic pessimism in analyst forecasts, but this bias is time-varying. A practical model should estimate the current level of analyst bias (e.g., using a rolling window of forecast errors) and adjust the weight on analyst revisions accordingly.

Implementation: Calculate rolling forecast errors (actual earnings minus consensus forecast) over a 12-month window. If analysts are currently pessimistic (negative average error), increase the weight on upward revisions. If optimistic (positive average error), decrease the weight.

8. Machine Learning Architecture

A practical ML model should:

  • Input features: SUE (standardized unexpected earnings), analyst revision magnitude, analyst revision direction, information uncertainty, analyst ability/coverage, forecast pessimism measure, text sentiment
  • Target: 20–60 day forward returns (to capture PEAD)
  • Architecture: Gradient boosting (XGBoost, LightGBM) or neural networks with attention mechanisms to capture non-linear interactions
  • Validation: Time-series cross-validation with walk-forward testing to avoid look-ahead bias
  • Subsampling: Separate models for high-IU and low-IU stocks, as the signal strength differs

Current Macro Context and Data Availability

The empirical evidence on earnings surprises and analyst revisions is robust across the 1998–2009 period covered by most papers. However, several macro and structural changes since 2009 warrant consideration:

  1. Increased Analyst Coverage and Efficiency: Sell-side analyst coverage has consolidated, and information dissemination has accelerated. The PEAD anomaly has weakened for numerical earnings surprises ([P8]), suggesting that markets are more efficient at processing traditional signals.

  2. Rise of Alternative Data and Sentiment: The frontier of predictability has shifted from numerical surprises to qualitative signals (text, sentiment, insider trading). Analyst revisions, which incorporate qualitative information, may be more valuable than in the past.

  3. Passive Investing and Reduced Price Discovery: The growth of passive investing may have reduced the efficiency of price discovery, potentially increasing the predictive power of earnings surprises and analyst revisions. However, this effect is not directly quantified in the available papers.

  4. Earnings Volatility: Post-2009, earnings volatility has increased (particularly during COVID-19 and subsequent recessions), which may amplify the predictive power of earnings surprises and analyst revisions.

Recommendation: Practitioners should validate the joint signal's predictive power on recent data (2015–present) before deploying in production. The evidence from 1998–2009 is strong, but market structure changes may have altered the signal's efficacy.

Quantitative Summary: Joint Effect vs. Individual Effects

Hypothetical Regression Framework

Based on the evidence, a practical regression model might take the form:

$$R_{i,t+1:t+60} = \alpha + \beta_1 \cdot \text{SUE}{i,t} + \beta_2 \cdot \text{RevisionChange}{i,t} + \beta_3 \cdot \text{IU}{i,t} + \beta_4 \cdot (\text{SUE}{i,t} \times \text{IU}{i,t}) + \beta_5 \cdot (\text{SUE}{i,t} \times \text{RevisionChange}{i,t}) + \epsilon{i,t}$$

Where:

  • $R_{i,t+1:t+60}$ = 60-day forward return
  • $\text{SUE}_{i,t}$ = standardized unexpected earnings
  • $\text{RevisionChange}_{i,t}$ = change in consensus recommendation or analyst forecast
  • $\text{IU}_{i,t}$ = information uncertainty (idiosyncratic volatility, analyst coverage, etc.)

Expected coefficient signs (based on evidence):

  • $\beta_1 > 0$: Positive earnings surprises predict positive returns (PEAD)
  • $\beta_2 > 0$: Upward revisions predict positive returns ([P1])
  • $\beta_3 < 0$: Higher information uncertainty is associated with larger drift (more underreaction initially, but larger subsequent drift)
  • $\beta_4 > 0$: The earnings surprise effect is amplified in high-IU environments ([P6])
  • $\beta_5 > 0$: The joint effect of earnings surprise and upward revision is multiplicative (both signals pointing in the same direction amplify the effect)

Empirical magnitudes (from [P6] and [P1]):

  • The PEAD effect for extreme earnings surprises is on the order of 5–10% abnormal return over 60 days in high-IU environments
  • The revision change effect is robust but smaller in magnitude, on the order of 2–5% abnormal return
  • The joint effect (surprise + revision in the same direction) is likely 7–15% abnormal return, suggesting multiplicative rather than purely additive effects

Dominance and Interaction

Which signal dominates?

  • In low-IU environments (high analyst coverage, transparent firms): Analyst revisions may dominate, as the earnings surprise is quickly priced in. The revision captures forward-looking sentiment shifts.
  • In high-IU environments (low analyst coverage, opaque firms): Earnings surprises dominate, as the market underreacts more severely. Analyst revisions amplify the effect.

Redundancy assessment:

  • The signals are not redundant ([P1] demonstrates orthogonality of revision changes to earnings surprises)
  • The signals are not purely additive (interaction terms are likely significant, particularly in high-IU environments)
  • The signals are partially reinforcing (positive earnings surprises are more likely to trigger upward revisions due to analyst forecast pessimism, [P9])

Conclusion

The empirical evidence demonstrates that earnings surprises and analyst revisions are complementary signals with distinct information content. Earnings surprises capture the market's initial underreaction to earnings news, with the magnitude of underreaction modulated by information uncertainty. Analyst revisions, particularly changes in consensus recommendations, contain information orthogonal to earnings surprises and predict returns across a broad range of stocks and market conditions.

The joint predictive power of the two signals is neither purely additive nor multiplicative, but rather conditional on market structure and firm characteristics. In high-information-uncertainty environments, the joint effect is amplified, with earnings surprises and upward revisions reinforcing each other to produce larger abnormal returns. In low-IU environments, analyst revisions may dominate, as they capture forward-looking sentiment shifts that the market has not yet fully incorporated.

For machine learning practitioners, the optimal approach is to treat earnings surprises and analyst revisions as separate features in a predictive model, with interaction terms to capture non-linear effects. The model should condition on information uncertainty and analyst forecast bias, and should weight analyst revisions by analyst ability. Text-based signals from earnings calls and analyst reports should be incorporated, as they contain information that numerical surprises miss, particularly in recent years as markets have become more efficient at processing traditional signals.

The evidence supports a 20–60 day holding period to capture post-earnings-announcement drift, with the strongest returns accruing to high-IU stocks with large earnings surprises and upward analyst revisions. Practitioners should validate these findings on recent data, as market structure changes since 2009 may have altered the signal's efficacy, though the fundamental mechanisms documented in the literature remain sound.