Post-Earnings Forecast: Predicting Post-Earnings Drift on the S&P 500
An end-to-end empirical report: data, models, and trading evaluation
This report walks through the full project end to end: what post-earnings drift is and why it matters, which stocks and years are covered, how earnings events look across sectors and fundamentals, how prices behave around the announcement, what earnings language reveals, how the prediction task is set up, how the models perform, and whether the signal survives as a trading strategy. Every chart and recomputed number is produced live from the vendored data/ subset when the document renders. The canonical hosted copy lives at https://richard-pears.dev/pead-report/.
1. Post-earnings-announcement drift
Four times a year, every listed company reports quarterly earnings: revenue, profits, guidance. In an efficient market the stock price should jump to the new fair value within minutes and then move randomly. It does not. Since Ball and Brown (1968) first documented it, and Bernard and Thomas (1989) quantified it at roughly 25% annualised abnormal return for a hedge portfolio, decades of research confirm the same pattern: after the announcement-day jump, prices keep drifting in the direction of the surprise for days and weeks. That slow continuation is post-earnings-announcement drift (PEAD).
Why it persists is itself the puzzle. The textbook explanations:
- Investor underreaction. Earnings releases are dense; analysts and investors anchor on prior beliefs and update too slowly, so information diffuses over days rather than minutes.
- Limits to arbitrage. Drift is concentrated in smaller, more volatile, harder-to-borrow names and around periods of high uncertainty, exactly when professional capital is most constrained.
- Behavioural frictions. Disposition effects, inattention, and post-announcement news flow (analyst revisions, media coverage) stretch the adjustment.
The classic predictor is standardised unexpected earnings (SUE): how far reported EPS lands from consensus, often combined with price momentum. But no single signal explains the drift fully: price action, fundamentals, management tone, news sentiment, and market regime each carry a fragment. That is the gap this project targets.
Research question: can a unified model over technical, fundamental, and language signals predict both the direction and the magnitude of the post-earnings drift well enough to trade, and if so, which signals actually do the work?
The investor decision (hold or exit after earnings) becomes a supervised task with two outputs per earnings event: a direction class (Low / Moderate / Strong peak move) and a magnitude (the peak return itself). The illustration below fixes the convention used everywhere in this report.
The one-day gap between the feature window and the target window is the project’s central discipline: nothing measured after the t+1 close may leak into the inputs. Entry is therefore always tradeable: the model only ever sees what an investor would have known at the entry price.
2. Sample and market coverage
The dataset covers S&P 500 constituents from January 2014 to May 2026: 503 distinct symbols and 24,048 quarterly earnings events, roughly one row per company per quarter. Coverage builds up over time as history accumulates and the constituent list evolves.
symbols: 503 | events: 24,048 | span: 2014-01-07 → 2026-05-14
2.1 Sector composition
Technology, Industrials, and Financial Services contribute the most events; Energy and Basic Materials the fewest, mirroring the index itself. Sector labels come from the company-info table joined on (symbol, earnings_date) at a 100% match rate (149 events carry no sector and are grouped as Unassigned).
sample: 135 companies across 12 sectors
Drift is not a Tech story alone. Mean peak returns sit in a tight band across sectors. The phenomenon is market-wide, which is why the models are trained pooled across all names rather than sector by sector.
| sector | events | mean_peak |
|:-----------------------|---------:|------------:|
| Technology | 3863 | 0.05 |
| Industrials | 3515 | 0.0421 |
| Financial Services | 3372 | 0.0414 |
| Healthcare | 2845 | 0.0405 |
| Consumer Cyclical | 2583 | 0.047 |
| Consumer Defensive | 1641 | 0.032 |
| Real Estate | 1521 | 0.0374 |
| Utilities | 1506 | 0.0316 |
| Communication Services | 1114 | 0.0475 |
| Energy | 1017 | 0.0548 |
| Basic Materials | 952 | 0.0488 |
| Unassigned | 119 | 0.032 |
3. Earnings outcomes and accounting fundamentals
Each event carries the reported EPS, the consensus estimate, the surprise, and a compact set of accounting fundamentals (growth, margins, leverage, cash generation, ROE; see docs/fundamental_indicators.md for definitions).
beat 73.2% | met 6.9% | miss 19.9% (n=24,294)
Nearly three in four events beat estimates, the well-known optimistic-bias ratchet, where consensus drifts down into the print. The key question is whether the size of the surprise predicts the size of the drift.
correlation(surprise %, peak return) = 0.0341
| group | mean | count |
|:--------|-------:|--------:|
| beat | 0.0427 | 17640 |
| met | 0.0377 | 1650 |
| miss | 0.0458 | 4758 |
Beats, meets, and misses drift to peaks within about a point of each other on average. This is the report’s first major finding and the reason a single-signal strategy fails: the surprise signs the direction of the announcement jump, but the drift that follows is decided elsewhere: volatility, positioning, and regime, as the later sections show.
The accounting backdrop is healthy large-cap America: profitable, cash generative, moderately levered:
| | median |
|:-------------------|---------:|
| revenue_growth_qoq | 0.0198 |
| gross_margin | 0.4351 |
| debt_to_equity | 1.6849 |
| fcf_margin | 0.1183 |
| roe | 0.037 |
4. Price behaviour around earnings announcements
4.1 Timing of the post-announcement peak
The target is the highest high over the nine trading days t+2…t+10. The peak is U-shaped: it lands most often on day 2 (immediate follow-through) or day 10 (slow burn into the window edge), and least often mid-window.
mean 6.1 days | median 6 | day-2 share 21.5% | day-10 share 22.4%
The shape matters for strategy design: exiting at a fixed horizon captures one hump or the other, while a take-profit rule can catch either. That is the logic behind the Strat-1/2/3 harness in §8.
4.2 Market volatility regime at entry
Market-wide volatility at entry is the single strongest observable companion of drift. Median VIX at t+1 is ~16.5; when the VIX runs hot, realised peaks run higher, compensation for holding through turbulence.
median VIX at entry: 16.5
4.3 Distribution of the drift target
Target peak returns average ≈4.3% (std ≈4.5%, median ≈3.2%), right-skewed with a fat right tail. Classes split Low <2% / Moderate 2–4% / Strong ≥4%, leaving a Strong-heavy mix (≈40%) and a majority-class baseline DA of 0.332.
mean 0.0430 | std 0.0455 | median 0.0318
5. Textual signals: earnings-call and news sentiment
Two language channels feed the models. FinBERT scores every earnings-call transcript (positive/negative probabilities, length-weighted per event); news headlines are scored into pre- and post-announcement windows with relevance weighting (sentiment_stg.py). Raw transcript text is stripped before modelling; only the scores travel forward.
mean pos 0.458 | mean neg 0.409 (n=24,286 scored; 8 events without transcript scores)
| window | mean |
|:-------------|-------:|
| Overall pre | 0.182 |
| Overall post | 0.1799 |
| Ticker pre | 0.1862 |
| Ticker post | 0.1842 |
A caution these charts already hint at, confirmed in §7: language is polite and backward-looking, news tone is flat across the announcement, and the target is an intraday peak. The models end up listening mostly to volatility; sentiment contributes almost nothing measurable.
6. Modelling setup
Each earnings event becomes one row of 258 feature columns: 18 TA-Lib technical indicators × 12 trading days (216: RSI, MACD, ROC, EMAs, ADX, ATR, Bollinger, volatility, OBV, VWAP, all normalised cross-stock) + 12 VIX + 2 FinBERT transcript scores + 4 news-sentiment scores + 12 fundamentals + 12 sector dummies. Imputer and scaler are fit on train folds only.
Validation is a 20-fold expanding walk-forward over test quarters 2021 Q1 → 2025 Q4 (TabNet; the DNN run covers 16 folds from 2022 Q1, 7,967 test rows vs 9,937). The training set grows each fold while the test set stays one quarter. That is the honest way to evaluate a market model, since no fold ever trains on its future.
train history 12,683 → 22,101 events; val/test ≈ 496/497 events per fold
The model ladder runs from isolation to contribution:
- B1–B4 ablation (RF/XGBoost): cumulative modalities: B1 technical (239 features), B2 + fundamentals/sector (250), B3 + transcript (254), B4 + news (258). Result: direction accuracy degrades monotonically B1 0.4505 → B4 0.4430 and pooled R² falls. Each added modality dilutes rather than helps (
docs/b1-b4_ablation.md). - Base-paper reproduction (Hajek et al. 2025 LSTM/BiLSTM/CNN-LSTM): adapted to the harder peak-return target, LSTM reaches RMSE ≈4.48 / DA ≈42.6%. Ranking is preserved vs the paper’s CAR target with an expected absolute gap (
src/modeling/basepaper/base_paper_reproduction.ipynb). - TabNet classifier + regressor per fold on a Modal T4 (
src/modeling/tabnet/modal_tabnet_pead.py; tuning check intabnet_pruned.ipynb: Optuna adds only 2–4pp, a structural ceiling). - Multitask DNN: shared encoder [256,128,64] + BN/ReLU/Dropout(0.3), joint loss
α·CE + (1−α)·MSE(multitask_dnn_updates.ipynb).
7. Predictive performance
Both learned models beat the 0.332 majority baseline by 12–13 points; the DNN edges TabNet on direction (0.457 vs 0.450) and error (RMSE 0.0428 vs 0.0446). Recomputed direction accuracy from the per-prediction tables matches the stored fold summaries to <0.001. One bookkeeping note: the strategy notebooks trade the earlier results.parquet trial (DA 0.4371); §8 replays that trial so the ledger matches, while this section compares the best available trials.
7.1 Error analysis
7.2 Predicted conviction as a risk ranking
Actual realised drift rises monotonically with predicted class under both models: a usable risk rank. That ordering, not the raw accuracy, is what the strategy monetises.
8. Economic evaluation: trading-strategy backtest
Rules (from TABNET_STRAT_4.ipynb / DNN_STRAT_4.ipynb): $50k capital, $200 per trade, enter long at the t+1 close when the predicted class is Moderate or Strong, exit at the first day the high hits entry·(1+T) else the t+10 close. Strat-1 fixes T=2%; Strat-2 sweeps T ∈ 1–5% by Sharpe (best: T=5.0%); Strat-3 split-sells half at T/2. This replay runs the notebook-selected T=5% live from the vendored tables, trading the same results.parquet trial the notebooks use and applying their filters (pre-2026 signals, complete t+2…t+10 windows), with unconstrained cash and no costs. As a checksum, the same code at T=2% reproduces the notebook ledger to the dollar (8,221 trades, +$6,637).
TabNet signals 8,221 | matched 8,221
TabNet T=5%: trades 8,221 | hit-rate 38.0% | total P&L $12,105 at $200/trade, no costs
DNN T=5%: trades 5,648 | hit-rate 42.0% | total P&L $7,861 at $200/trade, no costs
8.1 Threshold selection
The sweep covers T=1% to 5%. Total P&L rises monotonically for both models while the hit-rate more than halves; the simplified business-day Sharpe agrees with the notebook Sharpe that T=5% is best (notebook cash-constrained levels run higher: TabNet 2.33, DNN 1.77). The overlays show what the trade-off feels like: higher targets compound harder with deeper interim dents.
| model | T | pnl | hit | sharpe |
|:--------|------:|---------:|------:|---------:|
| TabNet | 0.01 | 2868.47 | 0.858 | 1.097 |
| TabNet | 0.015 | 4921.01 | 0.793 | 1.53 |
| TabNet | 0.02 | 6636.78 | 0.726 | 1.802 |
| TabNet | 0.025 | 7903.4 | 0.659 | 1.906 |
| TabNet | 0.03 | 8436.4 | 0.588 | 1.829 |
| TabNet | 0.035 | 9530.54 | 0.53 | 1.903 |
| TabNet | 0.04 | 10297.5 | 0.473 | 1.926 |
| TabNet | 0.045 | 11310.8 | 0.426 | 1.982 |
| TabNet | 0.05 | 12104.7 | 0.38 | 1.991 |
| DNN | 0.01 | 1059.81 | 0.864 | 0.491 |
| DNN | 0.015 | 2742.26 | 0.809 | 1.042 |
| DNN | 0.02 | 3898.07 | 0.748 | 1.303 |
| DNN | 0.025 | 4861.31 | 0.688 | 1.44 |
| DNN | 0.03 | 5144.89 | 0.62 | 1.348 |
| DNN | 0.035 | 5862.98 | 0.566 | 1.405 |
| DNN | 0.04 | 6472.22 | 0.51 | 1.446 |
| DNN | 0.045 | 7205.64 | 0.464 | 1.498 |
| DNN | 0.05 | 7861.12 | 0.42 | 1.525 |
8.2 Equity and drawdown
At the selected T=5%, both models compound through the sample with one shared drawdown around the 2022 macro-volatility episode.
TabNet final $12,105 | DNN final $7,861 | TabNet deepest drawdown $-3,201
8.3 Exit anatomy
At T=5%, three-fifths of TabNet exits never touch the target and ride to t+10; the rest bank +$10 within days. Strong signals convert at roughly twice the rate of Moderate ones, so conviction carries into execution.
| exit | n | share | pnl |
|:------------|-----:|--------:|---------:|
| take-profit | 3120 | 0.38 | 31200 |
| t+10 close | 5101 | 0.62 | -19095.3 |
8.4 Split-sell
Selling a fraction early (Strat 3) banks partial profit sooner. The 50/50 split compounds hardest here (+$10,004 live; notebook Sharpe 2.23); every split trails the no-split +$12,105 on P&L while smoothing the path.
no split: P&L=$12,105 win=60.7% Sharpe(simplified)=1.99
70/30: P&L=$10,303 win=74.3% Sharpe(simplified)=1.96
60/40: P&L=$9,904 win=75.2% Sharpe(simplified)=1.96
50/50: P&L=$10,004 win=76.3% Sharpe(simplified)=2.03
8.5 Stop-loss sweep
Adding any stop-loss below the 5% target earns less than no stop at all. The notebook’s Sharpe-selected -1% stop turns three-quarters of trades into -$2 scratches (quoted exit mix below) and halves total P&L to $5,594. The edge is in letting winners run to the target, not in cutting losers sooner.
no stop: P&L=$12,105 win=60.7%
-1%: P&L=$5,594 win=24.8%
-2%: P&L=$6,092 win=38.0%
-3%: P&L=$7,171 win=47.2%
-4%: P&L=$7,929 win=52.6%
-5%: P&L=$9,277 win=56.1%
quoted notebook exit mix at SL=-1%: 74.9% stopped (-$2 avg), 19.4% take-profit (+$10 avg), 5.8% ride to t+10
8.6 Year by year
| year | DNN | TabNet |
|-------:|------:|---------:|
| 2021 | nan | 2984 |
| 2022 | 2313 | 1662 |
| 2023 | 1584 | 1393 |
| 2024 | 2774 | 4552 |
| 2025 | 1214 | 1525 |
At T=5% the replay gives TabNet +$12,105 over 8,221 trades (hit 38.0%) and DNN +$7,861 over 5,648 (hit 42.0%), within tens of dollars of the cash-constrained notebook ledgers (+$12,138 over 8,221 with 13 skipped; +$7,774). Every full year 2021–2025 is profitable for both models. Losing quarters are scattered and shared across models, notably 2022 Q2: macro, not model. On deployed capital the notebooks report, at T=5%: +14.0% (2021), +7.0% (2022), +6.6% (2023), +19.9% (2024), +5.9% (2025).
9. Conclusions and limitations
- The drift is real but the surprise doesn’t price it. Beat rate 73%, yet beats and misses drift to the same peak on average and surprise size correlates ≈0 with drift size. One signal is not enough: confirmed.
- Volatility is the signal the models can actually hear. VIX regime at entry tracks realised drift monotonically; TabNet attention puts ≈92.5% on volatility/ATR features and <0.5% on fundamentals + sentiment combined. A daily/quarterly feature set predicting an intraday peak caps achievable R² near 0.12–0.14. The accuracy ceiling is structural, and tuning (+2–4pp) cannot lift it.
- A weak classifier can still be a strong risk ranker. Hit-rate and realised drift both rise monotonically with predicted conviction (62%→77%): the tradeable output.
- Economics beat statistics. Where classification fell short of the literature benchmark, the strategy validates: Sharpe 2.33, +$12,138 on $50k over ~5 years, profitable every full year.
Limits, stated plainly: zero transaction costs/slippage (a 0/5/10/25bp grid is open work); take-profit selection reuses its evaluation sweep (needs walk-forward threshold selection); Sharpe accounting must keep portfolio-level and deployed-capital figures separate; the risk layer is missing MaxDD/Calmar/Sortino, SPY and equal-weight benchmarks, and bootstrap intervals; news coverage is API-limited; the sample is current constituents only (survivorship); 2026 is a partial year and stays out of evaluation. The staged path is paper-trade → add frictions → small deployed allocation with oversight → quarterly retrain and monitor, not autonomous capital.
Appendix A. Where everything in the repo went
| Repo area | Disposition in this report |
|---|---|
src/ingestion/*.py, src/preprocessing/*.py (non-archive) |
§§2–5 describe; logic not rerun |
src/data/model_staging/tech_modeling_table.parquet |
§§2–4 live EDA |
src/data/model_staging/modeling_fundamentals.parquet |
§3 beats/misses/fundamentals |
src/data/model_staging/{nz_sentiment, finbert_tx_agg_weighted}.parquet |
§2 sectors, §5 sentiment |
src/data/model_staging/strat_table.parquet |
§8 live strategy replay |
src/notebooks/EDA_final.ipynb |
§§2–5 narrative source |
src/modeling/b1-5/b*.ipynb + docs/b1-b4_ablation.md |
§6 quoted ablation |
src/modeling/basepaper/base_paper_reproduction.ipynb |
§6 quoted baseline |
src/modeling/tabnet/modal_tabnet_pead.py, modal_tabnet_analysis_updated_results.ipynb, tabnet_pruned.ipynb |
§7 recomputed from vendored outputs |
src/modeling/dnn/multitask_dnn_updates.ipynb |
§7 recomputed from vendored outputs |
TABNET_STRAT_4.ipynb, DNN_STRAT_4.ipynb |
§8 quoted sweep + live replay |
All */archive/*, fundamental.ipynb, ppt/ms7_rework.ipynb, class notes/, temp/ |
excluded; see src/report/data/README.md |
tech_daily_table.parquet (594M), S3 backups, per-symbol FinBERT, transcript_staging/ |
excluded from the report subset |
Appendix B. Metric definitions
- DA: 3-class hit rate on the held-out quarter (Low <2% / Mod 2–4% / Strong ≥4% peak high over t+2…t+10). Baseline 0.332.
- MAE/RMSE: on decimal
target_return(×100 = pp); target mean ≈4.3%, std ≈4.5%. - Sharpe: annualised portfolio Sharpe on the daily equity curve reindexed to all trading days; deployed-capital and exit-day variants are separate diagnostics, not the headline.
- Win rate (strategy): fraction of entered trades hitting T before t+10 (distinct from classifier hit-rate).