Classification vs. regression in trading

Choose a model target that maps clearly to a decision, and measure whether it improves the trade process after costs.

The target changes the decision

Classification may estimate whether a setup reaches a defined outcome. Regression may estimate a future return, range, or expected value. Neither target is automatically more useful; the target should match the action the strategy can take.

Define the horizon, threshold, and treatment of overlapping trades before creating labels.

Evaluate the trading decision

A better prediction score does not necessarily mean a better strategy. Translate predictions into position selection or sizing, apply costs, and compare the result with a fixed-rule baseline. Look at calibration, coverage, turnover, and drawdown.

Do not optimise accuracy when the payoff is asymmetric.
Test threshold sensitivity.
Report how many trades the model filters out.
Measure the result after the actual execution policy.

Keep the output interpretable enough

The model output should have a clear role: accept, reject, rank, or size a known setup. A score that cannot be connected to an action makes risk control and incident review harder.

Use the simplest target and model that produces stable decision value out of sample.

Discuss this article

Ask a setup question, share a backtest, or compare notes with other algorithmic traders. Use a display name; your email remains private.

Create an account or sign in above to join the conversation.

Loading discussion…