Training Data Window
Definition
Training data window is the historical date range used to train an ML-augmented EA's model. Short windows (3-6 months) produce models tightly fit to recent regimes but vulnerable to regime shifts; long windows (3-5 years) produce regime-robust models but slower to adapt. The window choice materially affects model behaviour and decay rate.
In-depth: Training Data Window
Training data window selection is one of the most important hyperparameters in ML-augmented EA design. Different windows produce materially different model behaviour and have different implications for retraining cadence and edge longevity.
Training window selection considerations:
• **Recency relevance**: recent data reflects current microstructure (broker execution patterns, algorithmic competition levels, regime characteristics). Models trained on recent data perform well on current conditions • **Regime coverage**: longer windows cover more regime variations (trending, ranging, volatile, stressed). Models trained on diverse regimes generalise better when regime changes occur • **Sample size**: longer windows provide more training examples, reducing overfitting risk through statistical regularisation • **Non-stationarity**: financial time series are non-stationary; old data may reflect microstructure that no longer exists. Very old data can mislead model training rather than help • **Computation overhead**: longer windows produce larger training datasets, increasing training time and infrastructure cost
Typical training window choices by strategy class:
• **High-frequency scalping (M1-M5)**: 3-6 months. Recent microstructure dominates; older data is less relevant • **Day-trading (M15-H4)**: 12-24 months. Balances recency with regime coverage • **Swing-trading (H4-D1)**: 2-3 years. Slower strategies need more regime variation to validate • **Trend-following (H4+)**: 3-5 years. Trend identification benefits from extensive regime coverage • **Multi-strategy ensemble**: per-module windows; trend module on longer window, scalping module on shorter
Rolling vs static training windows:
• **Rolling windows**: training window slides forward over time; oldest data is dropped as new data is added. Most common for production EAs; combined with periodic retraining (weekly, monthly) • **Static windows**: training window fixed at deployment date; never updated. Suitable for academic research; inappropriate for production deployment where microstructure evolves • **Expanding windows**: training window starts at fixed past date and extends to current date; never drops old data. Useful for very long-horizon strategies but produces models that drift toward older patterns
Training window vs retraining cadence interaction:
• Short window + frequent retraining: model adapts quickly to changing conditions; risk of overfitting to noise • Long window + infrequent retraining: model is stable but slow to adapt; risk of stale microstructure assumptions • Optimal balance depends on strategy class and microstructure stability
For EA buyer evaluation: vendors using ML augmentation should document their training window choices and the rationale. Vendors who specify window length, sample size, and rolling-vs-static methodology signal serious engineering investment; vendors who just say "trained on historical data" are likely using ML as marketing without rigorous training discipline.