Why model.fit() Is the Least Interesting Line in an ML Trading System

The line model.fit() in an ML trading system may seem insufficiently interesting, but it hides the complexity of fitting the model together with the data puller, fold planner, and refusal check. These components have varying interpretations of the required historical data, and aligning them is a crucial task.

Within StratCraft, the model's role is to identify interactions that are challenging to encode manually. Momentum can vary based on liquidity, volatility, crowd participation, and broader market conditions. The model learns these interactions, and the exact returns returned by the model are not as precise as they appear due to financial data noise.

A common pitfall in ML trading research is treating exact returns as the primary target, such as predicting a stock's return for the next day. However, financial labels are inherently noisy, and a model's ability to separate stronger candidates from weaker ones is more valuable than exact return predictions.

Another shortcut is calculating numerous factors and feeding them into XGBoost for feature importance. While this approach yields a backtest that's easier to trust, it can lead to including invalid features, such as those not known at decision time, or multiple copies of the same exposure. The solution resides in the research layer, where validation geometry is treated as part of the data, not just a comment.

A critical aspect of successful ML trading systems is the embargo, which prevents overlapping observations from affecting training data. By adjusting the embargo, the model can account for information leakage and generate a more accurate out-of-sample test. In summary, the model.fit() line is just the tip of the iceberg in a complex ML trading system.