Why Past Performance Does Not Predict: Survivorship and Look-Ahead Bias

Reading-order note and boundary. Numbered #7 and drafted seventh, so it can reference every claim made earlier in the pillar. Boundary with two neighbouring articles: Survivorship Bias in Data: Why Dead Companies Matter covers the bias as it affects company and index data — this article covers performance records. And #8 covers biases a researcher introduces when testing a strategy; this article covers biases already present in the evidence before anyone analyses it.

It is not. It is a compressed statement of something much stronger: the historical record you are shown has usually been distorted before it reaches you, and even an undistorted record would predict far less than intuition suggests. Those are two separate problems — corrupted evidence, and weak evidence — and this article takes them in that order.

Three distortions in the record itself

1. Survivorship: the failures leave the sample. Funds that perform badly get closed or merged into others, and when they go, their records generally go with them. What remains in any database of "funds available today" is the set that survived — so the average performance of that set is not the average performance of the funds an investor could actually have chosen ten years ago. The distortion is not small, as the worked example quantifies: on plausible closure rates, an apparent average return can overstate the true experience by well over a percentage point a year, which compounds into a quarter of the outcome over twenty years. Every historical average of a group of funds is therefore an average of the winners, unless the source explicitly states that dead funds are included — and most do not.

2. Backfill and incubation: the record starts when it starts looking good. A firm can launch several funds quietly with small amounts of internal money, run them for a period, then market only the ones that did well — attaching their full history to the marketing material. The published record is accurate, in that those returns genuinely happened. It is also the top of a distribution the reader is never shown, because the funds that did poorly were closed before anyone outside the firm knew they existed. Related is the practice of a fund being added to a database along with its prior history, which means the database gains a track record that was selected for after the fact.

3. Look-ahead: using information that was not available at the time. This is the subtlest and the easiest to commit accidentally. Examples that recur. Restated financials — analysing a past decision using the corrected numbers, when the decision-maker had only the original, later-revised ones. Index membership — studying "the companies in the index today" over the last twenty years, which builds in the fact that they grew enough to qualify, and quietly excludes everything that fell out. Reporting delays — treating an annual figure as available on the last day of the financial year when it was published months later. And knowledge of what happened next, which is nearly impossible to unlearn: the analyst knows which crises came, which companies failed, and which sectors led, and that knowledge shapes what looks reasonable in hindsight. Look-ahead bias always flatters the past, because information you did not have is only ever useful.

And even a clean record predicts weakly

Suppose all three distortions were removed. The evidence would still be thinner than it looks, for reasons this pillar has already established. The detection problem. The alpha article shows that a modest genuine edge takes decades of record to distinguish from luck — 64 years at a one-point edge with typical variability. Records shorter than that cannot establish skill, however good they look. The luck baseline. The efficient-market article computes it: among a thousand managers with no skill at all, roughly thirty-one beat the market five years running and one manages a full decade. Impressive records exist in the absence of skill as a matter of arithmetic, not as a rare accident. Reversion. Extreme results in any noisy process tend to be followed by less extreme ones, because the extremity was partly noise and noise does not repeat. A period of strong performance is therefore weak evidence about the next period and is sometimes evidence about the last one. And the thing being measured changes. A record belongs to a manager, a strategy, and an asset size that may all have altered — funds grow, mandates shift, personnel leave, and a strategy that worked in a small fund can stop working in a large one. The record describes the past of something that may no longer exist in the same form.