2026-03-25

Backtested post-earnings announcement drift across 387 earnings events, 50 large-cap tickers, 2-year lookback (2024-04 to 2026-03). Spearman rank correlation of surprise % to forward return was significant at every horizon and strengthened over time: +0.159 at 1d, +0.151 at 5d, +0.185 at 10d, +0.156 at 30d, +0.174 at 60d, +0.201 at 90d. The drift persists; it doesn’t fade. That’s the rare shape. Most edges decay with horizon as noise dominates.

This backtest is from an options trading system I built that scores a portfolio of candidate features (momentum, sector flow, options skew, surprise magnitude). The question it answered was which of them survives a real out-of-sample test. PEAD did, decisively. Big beats (>5% surprise) drift +3.3% over 30 days at 60% win rate (n=167). The sweet spot tightens further with momentum-aware entry: stocks that had not already run up 5%+ pre-earnings, then beat: +3.5% avg, 63% WR, n=118. The pure surprise-magnitude curve scales monotonically: +16-20% surprise gives +6.8% avg / 75% WR (n=12). Most of the drift compresses into the first 5 days; entering on day 0-1 captures the most alpha, day 10 entry misses half the move.

Per-ticker reliability matters and isn’t uniform. CMCSA, GOOGL, INTU, MELI, NFLX show strong positive PEAD (correlation between +0.55 and +0.98 across 7-8 events each). JNJ (-0.86) and ABBV (-0.67) show anti-PEAD, where surprise predicts the opposite direction. The cross-sectional average is a fiction; any production system has to learn per-ticker reliability rather than apply the headline coefficient everywhere.

In a basket of features that mostly compete with each other on subtle in-sample fits, the one that holds up out-of-sample is the academic anomaly that has been studied for forty years and persists for behavioral reasons. Investors underreact to earnings information; the price correction plays out over weeks. Engineering edge, when it exists, looks more like leveraging a well-documented inefficiency than discovering a new one. PEAD is an effect to harvest, not a system to invent.