The fading of a strategy's excess return, in two senses. In trading, it is how fast a signal's predictive power fades after the signal appears, which sets how quickly a trade must be made and how long it is worth holding. In research, it is how an edge erodes over months and years as it becomes known, gets crowded or meets a new market regime. Both are read from a signal's measured skill: against the time since the signal in the first sense, against the calendar in the second.

Why does a strategy do worse live than in its backtest?

Usually for three reasons. The backtest was the best of many tries, so part of its return was luck that will not repeat. It left out costs such as spreads, market impact and borrowing fees, or assumed trades at prices that were not available. And the market changed after the test period: the edge was traded away or the regime shifted. The first two are errors in the research; only the third is decay in the strict sense.

How can you tell decay from an ordinary losing streak?

Often you cannot, at least not quickly. Assuming independent, normally distributed returns, the standard error of a Sharpe ratio estimated from T years is roughly the square root of (1 + SR²/2)/T. For a strategy with a true Sharpe ratio of 0.5 measured over four years, that is about 0.5, so four flat years are consistent with an intact edge. Sequential detectors help by adding up evidence as it arrives, and calibration checks show whether a model's stated odds still hold. A detector's threshold sets the trade-off between catching decay early and raising false alarms.

Does retraining a model fix strategy decay?

Only in part. Refitting on recent data helps when relationships have shifted, but it cannot restore an edge other investors have traded away, and refitting often on short windows can chase noise. A retrained model should be judged like any new one, on data it did not see. In Opulence Alpha the newest trained model becomes the champion automatically, but a separate gate on held-out data decides whether it carries trading weight, and a model that fails it is still scored every day.

FAQ · Strategy decay

Why do trading strategies stop working?

Reviewed against the platform's code on Sep 26, 2026

Strategy decay, or alpha decay, is the loss of a trading strategy's edge over time: the returns or forecasting skill it showed in research shrink, disappear or reverse once it is traded. Common causes are overfitting in the original research, other investors crowding into the same trade, changes in market regime and rising costs.

Why it matters

A strategy is chosen on its past, and the past flatters it. Part of a backtested edge is usually luck from the search that found it, so live results tend to start below the backtest even when something real was found. Real edges erode too. For 97 published return predictors, average returns were 26% below their in-sample level between the original sample's end and publication, and 58% below it after publication, a gap the authors attribute to investors trading on the findings (McLean & Pontiff, 2016). Crowded trades can unwind together, as many quantitative equity funds found in August 2007 (Khandani & Lo, 2011), and a relation that held in one regime can weaken or reverse in the next.

How it works

Decay is found by comparing live results with what the research led you to expect. A signal's information coefficient is tracked each period; a slide towards zero is often the first sign. Sequential change detectors such as CUSUM (Page, 1954) and the Page-Hinkley test add up small deviations from an expected level and raise an alarm once the running total crosses a threshold. Two-sample tests such as Kolmogorov-Smirnov ask whether recent inputs, forecasts or errors are distributed differently from before, and calibration checks ask whether stated probabilities still hold. A shift in the inputs alone is data drift; a change in how they relate to returns is real concept drift (Gama et al., 2014), the form decay takes.

How Opulence Alpha applies it

Opulence Alpha watches for decay model by model. As outcomes become known, a Page-Hinkley detector tracks each model's daily information coefficient at each forecast horizon, and a second tracks how much its top-decile picks churn. A model flagged as degraded loses the credibility its track record earned and must win back its weight in the blend on recent results. A finding that a model beats its baseline is set aside as inconclusive if a two-sample Kolmogorov-Smirnov test finds its daily metric shifted between the two halves of the test window. The 1- and 5-day models are scheduled to retrain weekly, every model monthly. The studies measure decay too: short-term reversal, their steadiest relation, has shrunk decade by decade since 1995.

Short-term reversal, smaller each decade →

Related concepts

Backtest overfittingBacktest overfitting is when a strategy is tuned to the noise of past data, so its backtest looks strong but its live performance decays. How to detect and avoid it.Information coefficientThe information coefficient (IC) measures how well a signal ranks stocks: its correlation with the returns that followed, across stocks each date, averaged.Market regime detectionMarket regime detection classifies the current market state — volatility, credit, liquidity, policy, the business cycle — so strategies and risk can adapt. How nine axes and a hidden semi-Markov model do it.Forecast calibrationForecast calibration checks whether stated probabilities come true as often as claimed: reliability diagrams, Brier score, calibration error and coverage.

All concepts →Sector behaviour in every market regime →

Questions

What is alpha decay?: The fading of a strategy's excess return, in two senses. In trading, it is how fast a signal's predictive power fades after the signal appears, which sets how quickly a trade must be made and how long it is worth holding. In research, it is how an edge erodes over months and years as it becomes known, gets crowded or meets a new market regime. Both are read from a signal's measured skill: against the time since the signal in the first sense, against the calendar in the second.
Why does a strategy do worse live than in its backtest?: Usually for three reasons. The backtest was the best of many tries, so part of its return was luck that will not repeat. It left out costs such as spreads, market impact and borrowing fees, or assumed trades at prices that were not available. And the market changed after the test period: the edge was traded away or the regime shifted. The first two are errors in the research; only the third is decay in the strict sense.
How can you tell decay from an ordinary losing streak?: Often you cannot, at least not quickly. Assuming independent, normally distributed returns, the standard error of a Sharpe ratio estimated from T years is roughly the square root of (1 + SR²/2)/T. For a strategy with a true Sharpe ratio of 0.5 measured over four years, that is about 0.5, so four flat years are consistent with an intact edge. Sequential detectors help by adding up evidence as it arrives, and calibration checks show whether a model's stated odds still hold. A detector's threshold sets the trade-off between catching decay early and raising false alarms.
Does retraining a model fix strategy decay?: Only in part. Refitting on recent data helps when relationships have shifted, but it cannot restore an edge other investors have traded away, and refitting often on short windows can chase noise. A retrained model should be judged like any new one, on data it did not see. In Opulence Alpha the newest trained model becomes the champion automatically, but a separate gate on held-out data decides whether it carries trading weight, and a model that fails it is still scored every day.

References

McLean, R. D. & Pontiff, J. (2016). Does Academic Research Destroy Stock Return Predictability? Journal of Finance 71(1), 5–32.
Khandani, A. E. & Lo, A. W. (2011). What Happened to the Quants in August 2007? Evidence from Factors and Transactions Data. Journal of Financial Markets 14(1), 1–46.
Page, E. S. (1954). Continuous Inspection Schemes. Biometrika 41(1/2), 100–115.
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M. & Bouchachia, A. (2014). A Survey on Concept Drift Adaptation. ACM Computing Surveys 46(4), 44.

Educational content about research methods. Not investment advice.