Machine Learning in Trading: Applications and Traps

Machine learning will find a signal in any data you give it — including pure noise. The discipline is making sure the signal is real.

ML has a seductive promise: feed it data, get a profitable model. In trading, that promise is half true. ML genuinely helps in some domains and confidently destroys capital in others. Knowing the difference is the whole job.

Where ML genuinely helps

1. Feature engineering at scale: discovering nonlinear interactions across hundreds of inputs that humans would never find
2. Cross-sectional ranking: predicting which stocks outperform peers this week, not the absolute return
3. Alternative data: extracting signal from satellite images, NLP on filings, sentiment from news
4. Risk modeling: forecasting volatility and correlations more accurately than GARCH
5. Execution: optimizing order slicing and routing in real time

Where ML reliably fails

1. Direct price prediction with raw OHLCV — the signal-to-noise ratio is near zero; you'll fit noise
2. Small samples with deep models — overfitting is essentially guaranteed
3. Regime changes — a model trained on 2010–2020 can break catastrophically in 2024
4. Non-stationary data — ML assumes the training distribution holds at inference; markets don't cooperate

The overfitting trap

With enough parameters, any model fits any historical series. A neural net can perfectly predict your training set and be worthless out-of-sample. Defenses:

• Purged walk-forward cross-validation: train on past, test on future, with a gap to prevent leakage
• Strict out-of-sample testing: hold out years of data the model never sees
• Simplicity bias: prefer linear models until you prove a nonlinear one adds value
• Regularization: L1/L2 penalties, dropout, ensembling
• Deflated Sharpe Ratio: correct for the number of strategies you tried

A realistic ML workflow

1. Frame the question carefully — predict ranking or direction, not raw price
2. Build clean features with economic meaning — don't dump raw OHLCV
3. Cross-validate with purged, time-aware splits (never random — that leaks the future)
4. Start simple: logistic regression, gradient boosting; only escalate to deep learning if needed
5. Measure out-of-sample with realistic transaction costs
6. Monitor for decay: rolling Sharpe; retrain or retire when it slips

Common mistakes

• Random train/test splits — these leak future into past. Always use chronological splits
• Hyperparameter tuning on the test set — kills the meaning of "out-of-sample"
• Survivorship and look-ahead in features — silently inflates results
• Treating ML as a black box — if you can't explain why it works, you can't predict when it stops working

Summary

ML is a tool, not a strategy. It shines for feature discovery, ranking, and risk modeling — and fails when asked to predict raw price from thin signals. Treat every model as a hypothesis: train it carefully, test it out-of-sample with realistic costs, and watch it like a hawk in production. The market doesn't care that your model has 99% training accuracy.