Post-Earnings Dip Backtest: 13,098 Events Say Don't Buy the Beat-and-Dip

A stock beats earnings estimates. The market sells it off 5% anyway. The obvious trade is to buy the dip. The fundamentals just proved out, so the sell-off is an overreaction. We ran this on US stocks from 2000 to 2025. The sell-off isn't an overreaction. It continues.

Contents

1. Method
2. What We Found
3. Dip Size Matters, But Not How You'd Expect
4. Why PEAD Beats Mean Reversion Here
5. The SQL Screen
6. Limitations
7. Takeaway

Method

Signal: epsActual > epsEstimated with ABS(epsEstimated) > 0.01. Then stock drops ≥ 5% from T-1 close to T+1 close (the two-day reaction window around the announcement). Both conditions must hold.

Universe: NYSE, NASDAQ, and AMEX-listed stocks with market cap above $1B USD. Data from FMP's earnings surprises endpoint, 2000–2025.

Dip categories: - dip_5: 5–10% sell-off (moderate sell-off) - dip_10: 10–20% sell-off (sharp sell-off) - dip_20: 20%+ sell-off (severe sell-off)

Return windows: Cumulative abnormal return (CAR) measured from T+1 close (dip bottom) at T+5, T+10, T+21, and T+63 trading days. CAR = stock return minus SPY return over the same window.

Winsorization: 1st/99th percentile to reduce outlier noise.

Data quality: A price-oscillation filter removes FMP phantom holiday rows and broken split adjustments before any return is computed. Events with a T+1 entry price below $1 are excluded, and single-window returns above 200% are treated as price artifacts and dropped.

Total events: 13,098

Data: FMP financial data warehouse, 2000–2025. Updated June 2026.

What We Found

The instinct is wrong. Stocks that beat earnings estimates but sell off 5% or more don't revert. They keep underperforming the market.

The overall T+21 CAR is -0.13% (t=-1.28). That's flat at 21 days, but by T+63 it deteriorates to -0.72% (t=-4.03). Not a recovery. The sell-off was telling you something.

p<0.05, p<0.01

The hit rate tells the same story. At T+63, only 46.3% of beat-and-dip stocks outperform the market. Coinflip odds would be 50%. You're worse than random.

Dip Size Matters, But Not How You'd Expect

The sharper the sell-off, the more it continues, except at the extreme.

Two things stand out.

The 10–20% sell-off group is the worst. These are stocks where the market sold off hard enough to signal something fundamentally wrong, but not hard enough to be "priced in." At T+21, they're down -0.58% vs SPY (t=-2.72). At T+63, -0.86% (t=-2.17). Both statistically significant.

The 20%+ group is different. The T+21 number is +0.27% and T+63 is +1.09%, but neither is statistically significant (t=0.47 and t=1.01). With n=684, the positive mean is being driven by a fat tail of recoveries, not a consistent effect. The median at T+63 is actually -1.36%. Half the severe dip stocks continue falling.

So: moderate dips are reliably bad, sharp dips are reliably worse, and severe dips are inconsistent with a slightly positive mean masking a negative median.

Why PEAD Beats Mean Reversion Here

Post-earnings announcement drift (PEAD) says stocks continue moving in the direction of the earnings surprise. Beat → drift higher. Miss → drift lower. The beat-and-dip pattern tests the opposite: beat → market says no → does the market eventually agree with the beat?

The data says the market wins. When a company beats estimates but the stock sells off 5%+, the market is likely pricing in something that's not in the EPS number. Guidance cuts. Revenue miss. Margin compression. One-time items. EPS can beat while every forward-looking metric looks worse.

The academic literature on PEAD (Ball & Brown 1968, Foster, Olsen & Shevlin 1984) focuses on the alignment between surprise direction and return. The beat-and-dip is a case where the two conflict: the number was good but the reaction was bad. The reaction contains more information.

The SQL Screen

This query finds current beat-and-dip candidates: stocks that just beat earnings but sold off 5% or more:

WITH recent_beats AS (
    SELECT e.symbol,
           CAST(e.date AS DATE) AS event_date,
           ROUND((e.epsActual - e.epsEstimated) / ABS(NULLIF(e.epsEstimated, 0)) * 100, 1) AS surprise_pct,
           ROUND(k.marketCap / 1e9, 2) AS mktcap_bn,
           p.companyName AS company_name,
           p.exchange
    FROM earnings_surprises e
    JOIN profile p ON e.symbol = p.symbol
    JOIN key_metrics k ON e.symbol = k.symbol AND k.period = 'FY'
    WHERE e.epsEstimated IS NOT NULL
      AND ABS(e.epsEstimated) > 0.01
      AND e.epsActual > e.epsEstimated
      AND CAST(e.date AS DATE) >= CURRENT_DATE - INTERVAL 30 DAY
      AND k.marketCap > 1000000000
      AND p.exchange IN ('NYSE', 'NASDAQ', 'AMEX')
    QUALIFY ROW_NUMBER() OVER (PARTITION BY e.symbol ORDER BY k.date DESC) = 1
)
SELECT rb.symbol,
       rb.event_date,
       rb.surprise_pct,
       rb.mktcap_bn,
       rb.company_name,
       -- Compare T+1 close to T-1 close for the reaction
       ROUND((p2.adjClose - p1.adjClose) / p1.adjClose * 100, 2) AS reaction_pct
FROM recent_beats rb
JOIN stock_eod p1 ON rb.symbol = p1.symbol
    AND CAST(p1.date AS DATE) = rb.event_date - INTERVAL 1 DAY
JOIN stock_eod p2 ON rb.symbol = p2.symbol
    AND CAST(p2.date AS DATE) = rb.event_date + INTERVAL 1 DAY
WHERE (p2.adjClose - p1.adjClose) / p1.adjClose <= -0.05
ORDER BY rb.event_date DESC, rb.surprise_pct DESC
LIMIT 30

The reaction_pct filter at -0.05 catches the 5%+ sell-off. This is a screen, not an entry signal. The backtest says these stocks underperform on average.

Run this screen live on Ceta Research → (pre-loaded query, no account required)

Limitations

Why EPS beats can mislead: A company can beat EPS via cost cuts, buybacks, or one-time tax benefits while its core business deteriorates. The market often sees through headline EPS. The stock reaction is frequently the more honest signal.

Market cap floor: The $1B floor focuses on liquid mid-to-large caps. Small caps may show a different pattern, but execution costs make the trade less practical there.

Survivorship bias: Companies that delisted due to failure are underrepresented, which slightly biases CAR estimates upward. The true result is likely a bit worse.

T+5 is flat, then it deteriorates. The very short-term recovery instinct isn't entirely wrong. T+5 is essentially zero (-0.04%, not significant). The problem is the 1-3 month drift. If you're trading with very tight exits, the signal might look different. But by the time most investors reconsider, it's T+63 territory.

Takeaway

"Buy the beat-and-dip" is the kind of trade that feels logical but fails empirically. The market's negative reaction to a genuine earnings beat usually means something the headline number doesn't show. 13,098 events over 25 years say: when the market says no after a beat, trust the market.

Taiwan is the notable exception, the only market in our study where beat-and-dip events reliably revert. India shows a positive direction at T+63 but no statistical significance. Both are discussed separately.

Part of a series: Post-earnings dip mean reversion tested across 11 exchanges. See Taiwan, India, and the global comparison.

Data: FMP earnings surprises + adjusted prices, 2000–2025. NYSE, NASDAQ, AMEX. Market cap > $1B USD. 13,098 beat-and-dip events. CAR vs SPY, measured from T+1 close.

Past performance does not guarantee future results. This is educational content, not investment advice.