After the Beat

1The question

When a company reports earnings above what analysts expected, its share price usually jumps on the day. An older finding in finance, known as post-earnings-announcement drift, is that the price often keeps moving in the same direction for weeks afterwards, as the market slowly absorbs the news.

We asked whether that drift still exists in the largest US companies, which are the most closely watched. We also asked whether a rule simple enough to follow by hand could capture it after costs.

2The rule

All four conditions are checked once, at the close of the reaction day. That is the report day for companies reporting before the market opens, or the next trading day for those reporting after the close.

3Data and method

• Earnings. Estimates, actual EPS and report times from Yahoo Finance's earnings history, for about 500 stocks from 2002 to 2026. That is 42,698 reports.
• Universe, 2016–2026. A point-in-time S&P 500: the companies that were in the index on each date, including ones that later left, where price data still exists. Market value is measured on the signal date.
• Universe, 2006–2015 (the blind test). Share counts before 2014 are unreliable, so the size tiers are set by rank instead: the 60 largest companies for the large tier and the next 170 for the mid tier. The rule's thresholds were chosen on 2016–2026 data, so this decade is out of sample.
• Prices. Daily prices, split-adjusted, without dividends, for both the strategy and SPY.
• Account. Starts at $100,000 with equal weights, rebalanced at each entry and exit, and a 0.05% cost on every trade. Cash earns nothing.
• Statistics. The "edge" is a trade's return above SPY, minus the average for all quarters of similar companies, so it is not just credit for owning stocks. The 95% ranges come from bootstrap resampling.

4Results

The 2006–15 range for all signals reflects two versions of the same simulation that differ in rounding and in when rebalancing happens.

A falling market: 2022

Technology shares fell about a third in 2022. Nine technology companies beat earnings by enough to pass the size and beat rules. Bought anyway, they would have averaged −11%, and only a third would have made money. The trend condition rejected all nine, because each one's MACD was below zero. Across all sectors that year, 79 beats passed the size and beat rules. The 36 that also passed the trend test averaged −3.3%, against −4.4% for all 79. The filter softened the year; it did not avoid it.

Where the return comes from

• The advantage over SPY builds slowly through the holding period: about +0.5 points after a month, +1.6 after three months, and +1.75 by the next report.
• Most of each trade's return is the market's own rise. Hedging out the market leaves the +1.75 points per trade.
• Losing trades cluster in market-wide shocks (early 2020, late 2021–2022, early 2025) rather than in any stock trait we could identify at entry.

5The blind test, 2006–2015

Out of sample, the large tier kept a small advantage of about +0.7 points per quarter over its baseline. The mid tier's 35% rule did not hold up (−0.4), which is why it is labelled experimental. Across both tiers the edge was +0.3 points, with a 95% range of −0.4 to +1.0, which is not distinguishable from zero.

The account still beat SPY over the decade, compounding about 8.5% a year against 5.0%. Much of that came from owning large companies through the 2009–2015 recovery. Survivorship bias is also worse in this decade (§8). A realistic expectation from this evidence is a modest advantage, not a repeat of 2016–2026.

6Sectors

We tagged every company with its GICS sector and repeated the test. Pharmaceutical and biotechnology companies are shown separately from the rest of health care.

• Technology was the only sector clearly above zero in 2016–2026: 114 trades, 75% rose, an average of +11.3% per trade and an edge of +4.9. Much of it came from AI and semiconductor names after 2023. In 2006–2015 the technology edge was close to zero.
• Pharmaceuticals and biotech showed no drift in either decade, and pharma beats moved less even on the day (+1.4% against +2.3% for technology). A plausible reading is that pipeline and trial news matter more to these share prices than quarterly earnings.
• Utilities were consistently weak, and other health care (devices, insurers) was weak in 2016–2026.
• Dropping the weak sectors changed the overall edge very little: 2016–2026 rose from +2.3 to +2.7 points, and 2006–2015 did not change. Choosing sectors from the same data would be overfitting, so the rule has no sector filter.

A focused variant

Many people prefer to hold a few positions rather than a dozen. Restricting the same rule to technology and the internet platforms (GOOGL, META, AMZN, NFLX) cut activity to about 15 trades a year and roughly four positions at a time. It produced 12.0% a year in 2006–2015 and 32.8% in 2016–2026, with smaller worst falls than the full rule.

The costs are concentration and hindsight. One bad quarter in a four-stock account matters far more, and the platform list names companies we already know succeeded. The BeatDrift app tracks this variant as a second paper portfolio so its live record can be compared fairly from now on.

7What did not work

Each of these was tested against simply holding to the next report, and none beat it:

• Stops and targets. Every profit target and stop-loss lowered the total. The least costly, a −20% stop, still lost about a sixth of the profit, though it capped the worst trade.
• Trend-following exits. SuperTrend on daily and hourly charts, and in-and-out trading inside the window, did worse than holding.
• Chart patterns. Harmonic-pattern entries inside the holding window did no better than the same patterns anywhere else.
• A bear-market filter (S&P 500 below its 200-day average). It removed the strong rebounds after 2020 and 2025.
• Machine learning. A neural network and gradient boosting, trained walk-forward over 2009–2026, ranked trades no better than a one-line rule based on the company's past beats.
• Shorting misses. Shorting after a miss lost money outright.

8Limitations

• Survivorship. Companies that left the S&P 500 are included only where price data still exists, which is about 40% of them. The 2006–2015 sample is missing many 2008 failures. Both decades are therefore biased upwards to some degree.
• No dot-com era. The earnings data begins in 2002, so the study cannot show how the rule behaved through the late-1990s boom and the 2000–2002 crash. The closest episodes covered are 2008, 2020 and 2022.
• The outlook condition is untested historically. The source keeps only final estimates, so a backtest of "next quarter's estimate is higher" would use information not available at the time. It is applied live only.
• Rebalancing. The simulated account rebalances at every entry and exit, about 100 times a year. That is more than most people would do, and it may overstate a hand-followed result.
• No dividends on either side, and a flat 0.05% cost. Taxes, spreads on less liquid days and currency effects are not modelled.
• Recent results dominate. A large share of the 2016–2026 compounding came from 2024–2026 and from a handful of AI-related stocks.
• Small samples. Several sector results rest on fewer than 50 trades, and their ranges are wide.

9The live test

Since 24 July 2026 the rule has run live: an automated scan of 524 large US stocks raises each signal, and a paper portfolio follows it with the same equal-weight method. By 9 October 2026 the paper account was up +9.9% against +6.3% for SPY, with 11 positions open. Eleven weeks is far too short to tell skill from luck. The point of the live test is to build a record that was never fitted to the past.

AAppendix: parameters and sources

Sources: Yahoo Finance earnings history and daily prices; historical S&P 500 membership (public constituent-change records with curated renames); TradingView for live market data and sector classifications; GICS sectors from the published S&P 500 constituent list. Analysis scripts: tvpet/earnings and tvpet/pit in the project repository.