Where machine learning helps systematic trading — and where it is a label

Almost every trading product now calls itself AI-powered. I want to be precise about my own work: the systems X-STOX ships today are rule-based. Oscillators, momentum and breakout logic, recovery grids, non-repainting signal rules. Each product page describes the actual mechanism.

That is not a complaint about machine learning. It is a statement about evidence.

The question worth asking

The useful question is not "can a model predict price?". It is narrower: given a rule-based system that already works, can a learned component make a specific decision inside it better than the rule does, out of sample, after costs?

Framed that way, there are candidates worth testing:

• Regime classification — deciding when a strategy's conditions are present at all.
• Execution decisions — estimating slippage or spread behaviour around events.
• Filtering — learning which of a rule's signals historically failed under similar conditions.

Why this is hard to do honestly

Financial data is noisy, regimes shift, and a model with enough freedom will always find a pattern in the past. The tests that matter are boring: hold-out periods the model never saw, walk-forward evaluation, realistic costs, and a willingness to publish "no improvement" as a result.

What I will publish here

When I write about this work, I will say what was tested, on what data, and what happened — including when the answer is "the rule was better". That seems more useful than another product called AI.