When not to use machine learning

Sometimes a clear rule, better data, or better execution solves the problem more reliably than a model.

Do not use complexity to hide uncertainty

If the strategy question is not defined, an ML model can produce an impressive output without answering a useful question. Start with a transparent baseline and identify the specific limitation that a model might address.

More parameters, features, and training time do not create more information in the market.

Check the simpler fixes first

A data timestamp bug, incorrect spread model, duplicate signal, or poor risk rule can look like a prediction problem. Fix data and execution quality before adding a model that may absorb the defect.

Confirm the baseline is implemented correctly.
Check whether a deterministic filter solves the stated problem.
Measure sample size and label stability.
Assess whether the team can monitor and retrain the model safely.

A model is a liability when it is unobservable

If you cannot explain feature availability, prediction freshness, failure behaviour, and decision thresholds, the model adds operational risk. It is better to run a simpler process you can test and monitor than a sophisticated one you cannot trust.

Use ML when the evidence and infrastructure justify it, not because a complex method sounds more advanced.

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…