Features Do Not Last Forever: Monitoring Predictive Power Over Time
Tracks individual feature predictive power over time, identifying 2 features with decaying alpha for potential removal.
Feature Alpha Has a Half-Life
Every feature in a trading model captures some market inefficiency. Over time, as more participants discover and exploit the same inefficiency, the feature's predictive power decays. S39 tracks each of V7's 38 features individually, monitoring their contribution to model accuracy over rolling 500-trade windows.
The tracking uses a combination of SHAP importance and directional accuracy. SHAP tells you how much the model relies on the feature. Directional accuracy tells you whether the feature is actually pointing the right way. A feature with high SHAP importance but declining directional accuracy is a ticking time bomb.
The Two Decaying Features
Over the 7.5-year backtest, S39 identified two features with meaningful alpha decay: bid_ask_imbalance and a custom momentum divergence metric. Both showed stable performance through 2020, then gradual decline through 2022-2024.
The bid_ask_imbalance decay is consistent with market microstructure evolution. As more algorithmic traders entered the market, the information content of the bid-ask imbalance decreased because the imbalance itself became a traded signal. The momentum divergence decay was more puzzling and likely related to changing market structure in the post-COVID era.
Decay Detection as Model Insurance
S39 is not about removing features immediately. It is about knowing which features to watch. The two decaying features still contribute positively to model performance. Their SHAP importance has dropped but is not zero. Removing them today would likely reduce performance slightly. S39 provides the monitoring framework to know when removal becomes the right call. If a feature's directional accuracy drops below 50% (worse than random), S39 flags it for immediate review. If it stays between 50-55% for an extended period, it flags for review at the next scheduled retraining cycle. This graduated response prevents both premature removal and delayed response to genuine decay.