Post-Earnings Announcement Drift

Thesis (edge)

A company reports earnings. The number is much better than analysts expected. The stock jumps 6 percent that day. In a perfectly efficient market, that would be the end of it. The new information is now in the price, and from that moment the stock is a fair bet again.

That is not what happens. Historically, the stock keeps drifting in the same direction for weeks or even months afterwards. The good news gets priced in slowly rather than instantly. The same is true in reverse for a bad report: the stock falls on the day, and then keeps sagging.

This is post-earnings announcement drift, and it is one of the oldest and most thoroughly documented anomalies in finance. It has been studied continuously since the 1960s, replicated across countries and decades, and it remains one of the most direct challenges to the idea that markets instantly reflect all public information.

The usual explanation is that investors underreact. They anchor on their previous view of the company and update their beliefs gradually. Analysts revise their forecasts in small steps over subsequent weeks rather than all at once. Slower institutional buyers take time to build positions. All of this leaves a residual drift that a systematic strategy can attempt to harvest.

Where it works (regimes)

The drift has historically been strongest in smaller, less followed companies, where information travels slowly and fewer analysts are watching. It is weakest in the mega-cap names that dozens of institutions cover in real time.

That fact is both the strategy's edge and its central problem. The place where the drift is largest is the place where trading costs are highest, and the two tend to cancel out. This is not an accident. It is the reason the anomaly has survived: it is genuinely hard to arbitrage away in exactly the names where it is strongest.

The effect has also visibly weakened over time as more systematic capital has hunted for it. Any honest treatment must say this plainly. The drift that academic papers documented in the 1980s is substantially larger than the drift available today, and a backtest run over the full history will materially overstate what is achievable now.

Signals

• Earnings surprise: the difference between what the company reported and what analysts expected. This must be scaled, because a 2 cent beat means something very different for a stable utility than for a volatile technology company. The standard approach divides the surprise by the historical variability of that company's surprises, which puts every stock on the same scale.
• Announcement return: how the stock actually moved on the day of the report. This is often a better signal than the earnings surprise itself, because it captures everything the market learned, including guidance, tone, and details in the report that no consensus number reflects. A company can beat on earnings and fall 8 percent because the outlook was terrible. The raw surprise says buy. The market reaction says otherwise, and the market reaction is usually right.
• Revision momentum: watch whether analysts are raising or lowering their forecasts in the days after the report. Sustained upward revisions reinforce the drift.
• Attention and volume: drift tends to be stronger when the announcement was under-noticed, which is one reason it concentrates in smaller names.

Portfolio construction

Rank stocks each week by surprise score. Take the top group long and the bottom group short. Hold each position for a fixed window, commonly 30 to 60 trading days, then exit.

Because companies report on a staggered schedule, the portfolio naturally becomes a rolling set of overlapping positions entering and leaving at different times. That is a feature: it spreads the trading over time rather than concentrating it, and it smooths the return.

Neutralize the exposures. If technology companies happen to have a great reporting season, an unhedged long book of positive surprises becomes a bet on technology, and your returns will be explained by that rather than by the drift. Hedge market beta, and neutralize sector and size, so that what remains is the drift itself.

Apply a liquidity floor before anything else, then honestly test whether the edge survives in the universe that clears it. Many implementations discover that the strategy works beautifully in stocks they cannot actually trade.

Risk model

The main risk is not a single dramatic loss. It is slow erosion. The drift is a small, diffuse edge spread across many positions, and it can be quietly consumed by transaction costs, by crowding, or simply by decaying to nothing.

Position-level risk is real too. A stock in the long book can report well, drift up for two weeks, then issue a profit warning and give back everything. Diversification across many names is the only defence, and it needs to be genuine diversification, not fifty stocks that are all in the same sector because that sector had a good quarter.

There is also a crowding risk that is specific to well-known anomalies. When many systematic funds run the same signal on the same universe with the same holding period, they buy the same stocks on the same day. That compresses the edge in normal times and can produce a sharp unwind if they all reduce risk together.

Costs & implementation

This is a moderate-to-high turnover strategy. Positions are held for weeks, and the universe is broad, so the trading bill is substantial. Costs are the single biggest determinant of whether a live implementation matches its backtest.

The short leg is a particular practical challenge. The negative surprises with the strongest drift are frequently small, beaten-down names that are expensive to borrow or unavailable entirely. A backtest that assumes free and unlimited shorting is not describing a real strategy. Many practitioners end up running the long side only, which halves the theoretical edge and adds market exposure that must then be hedged.

The data requirements are strict. You need point-in-time consensus estimates, meaning what analysts actually expected before the report, not a database that has quietly been updated since. And you need exact announcement timing. A stock that reports after the close on Tuesday cannot be traded at Tuesday's close, and a backtest that does so will produce spectacular and entirely fictional returns.

Failure modes

• Look-ahead bias in the consensus data. This is the most common and most damaging error in the entire strategy.
• Getting announcement timing wrong and capturing the announcement-day jump itself, which is not available to you and which dominates the measured return.
• Ignoring the announcement-day reaction and trading purely on the accounting surprise, which misses most of the information.
• Assuming the historical effect size still applies. It does not, and the decay is well documented.
• Building the strategy in illiquid names where it works and then being unable to trade it.
• Failing to neutralize sector exposure and mistaking a sector bet for an anomaly.

Our Notes & Suggestions

This is one of the best strategies to study early, because it teaches the difference between an effect that is real and an effect that is tradeable. Post-earnings drift is unambiguously real. It has been confirmed for sixty years across dozens of markets. And it is still difficult to make money from, because the costs, the shorting constraints and the decay have eaten most of what was once a very large edge.

If you build it, build the honest version. Point-in-time data, correct timestamps, realistic costs, a liquidity filter applied before the backtest rather than after, and a separate result for each decade so you can see the decay with your own eyes.

The most useful practical refinement is to lean on the announcement-day reaction rather than the raw earnings surprise. Let the market tell you what the report meant. A company that beats and rallies is a very different animal from one that beats and falls, and the accounting number alone cannot tell them apart.