Signal Decay: Why Published Edges Stop Working

A predictive signal is not a permanent property of markets. It is a description of a relationship that held over a period, and its usefulness depends on how many other participants are acting on the same relationship.

That sounds obvious stated plainly, and it is routinely ignored in practice. Strategies get evaluated on how well they worked historically, with no allowance for the fact that the historical period is the one in which nobody else was using them.

There is a substantial body of evidence measuring exactly how much predictive power survives once a signal becomes public knowledge, and the numbers are specific enough to plan around.

Where This Bites for Model-Driven Approaches

Any AI investment strategy built on publicly available data inherits this problem directly, and arguably in a sharper form.

The reason is structural. Machine learning applied to market data searches for relationships in a dataset that thousands of other participants also hold, using techniques documented in publicly available literature and increasingly available in off-the-shelf libraries.

Which means the relevant questions are not about model quality:

A model that finds a real relationship in public data has found something others can find too.

What the Publication Study Found

The canonical measurement examined what happened to documented return predictors after they appeared in academic journals.

The study covered 97 variables shown to predict cross-sectional stock returns and found that portfolio returns are 26% lower out-of-sample and 58% lower post-publication, with the out-of-sample decline serving as an upper bound estimate of data mining effects, implying a 32% lower return attributable to publication-informed trading.

Separating Two Causes

That decomposition is the important part. Some of the decline is statistical: a relationship discovered by searching a dataset will look weaker on data it was not fitted to, regardless of whether anyone trades it.

The rest is behavioural. Once published, participants act on the finding, and their trading moves prices toward where the signal said they should be, which removes the opportunity that generated the returns.

Where Decay Was Largest

Two further findings from the same work sharpen the picture. Post-publication declines were greater for predictors with higher in-sample returns, meaning the most impressive-looking signals decayed the most. And returns after publication remained higher in portfolios concentrated in stocks with high idiosyncratic risk and low liquidity, which are precisely the stocks that are expensive to trade.

The survivors are the signals that are hard to exploit, not the ones that are hard to find.

Crowding as the Mechanism

More recent work has focused on the process by which decay occurs, which is more useful than the endpoint.

The same publication study found that predictor portfolios exhibit post-publication increases in correlations with other published-predictor portfolios. Research on factor crowding has since built on this, examining how systematic strategies decay and how crowding can be modelled and measured across factors, with the recognition that factors do not crowd equally.

The practical reading is that decay is not uniform. Some signals attract capital quickly and stop working; others persist because the conditions for trading them remain unattractive.

Why Model-Driven Approaches Compress the Timeline

The publication study measured decay over a period when discovering a predictor required a research team, a journal cycle and manual implementation. Several parts of that pipeline have shortened considerably.

Standard datasets are widely licensed, modelling techniques are documented and implemented in shared libraries, and computational cost has fallen. A relationship that once took years to move from discovery to widespread use can now travel that distance far faster.

None of this makes systematic approaches unworkable. It changes the expected lifespan of any given signal, which is a planning question rather than a viability question.

What Tends to Survive

Given the evidence, some categories hold up better than others:

The fourth category is the most durable, because compensation for bearing risk does not disappear when more people learn about it.

What This Means in Practice

The realistic expectation for anyone building or buying a systematic approach is that documented performance overstates what follows, by an amount the evidence puts at roughly half.

That has two implications worth acting on. Backtested results need discounting before they are compared against alternatives. And any strategy needs monitoring for decay rather than being assumed stable, with a defined process for deciding when a signal has stopped working.

Neither implication requires abandoning systematic methods. Both require treating any specific edge as something with a shelf life, which is a different posture from treating it as a discovery.