Machine learning for trading: a realistic starting point

A useful first ML task might rank candidate trades, estimate a probability under a fixed setup, or identify when a known strategy should reduce activity. Avoid starting with the vague goal of predicting price; define the decision, horizon, and action before selecting a model.

The model should improve a measurable decision compared with a simple non-ML baseline.

Trading data is sequential. Randomly splitting rows can place information from the future in the training set while evaluating earlier observations. Use time-aware splits, fit transformations on training data only, and preserve the order of decisions.

An ML score should not bypass position limits, drawdown controls, or execution checks. A model can be uncertain, stale, or out of distribution. Use it as one input to a controlled process, not as permission to increase leverage.

The simplest model that adds stable out-of-sample value is usually a better first step than a complex model that cannot be explained.