Detecting Strategy Decay Before It Costs Real Money
60-day rolling Sharpe ratio with decay detection alerts that flags significant deviations from historical baseline performance.
Why Rolling Sharpe Over Cumulative
Cumulative Sharpe ratio tells you how the strategy has performed overall. It cannot tell you how it is performing now. A strategy that generated most of its returns in 2020 and has been flat since 2024 will still show a decent cumulative Sharpe. Rolling Sharpe exposes the truth.
S26 computes a 60-day rolling Sharpe ratio and compares it to the historical baseline. When rolling Sharpe drops below 50% of the full-sample Sharpe for 20 consecutive days, S26 generates a decay alert. This threshold was calibrated to avoid false alarms during normal drawdown periods while catching genuine performance degradation.
Alert Thresholds and False Positives
The 50% threshold and 20-day persistence requirement were tuned using historical data. During the normal backtest period, the system generated 3 false decay alerts, all during expected drawdown periods that subsequently recovered. Adjusting to a 40% threshold eliminated false alarms but would have missed one genuine degradation event in a synthetic test.
The 50%/20-day combination provides the best balance between sensitivity and specificity. In live trading, a decay alert triggers a review process: check feature importance (S11), validate model calibration (S37), and run fresh Monte Carlo simulations (S41). The alert does not automatically stop trading because temporary underperformance is normal.
The Monitoring Philosophy
S26 embodies the principle that deployment is not the end of model development. A frozen model with frozen weights will eventually decay as market dynamics evolve. The question is when, and how you detect it. S26 provides the earliest quantitative signal that the strategy's edge may be eroding. Combined with S11 (feature decay) and S39 (feature predictive power), it creates a monitoring triangle that covers model-level, feature-level, and return-level degradation detection. Catching decay at the rolling Sharpe level means you have months of warning before cumulative performance is materially affected.