AI in trading — what is actually working in 2026
The most useful thing we can say about AI in trading in 2026 is that the public conversation is roughly two orders of magnitude noisier than the actual practitioner experience. This is a long note on what has earned a place in our production stack, what has not, and why the gap between the marketing narrative and the operational reality is so persistent.
- ML has earned a defined, narrow role: feature extraction, calibration, and regime classification.
- End-to-end deep models that emit positions from raw price have not survived our walk-forward tests.
- Large language models are useful in research workflow — summarising filings, extracting structured data — but do not generate signals in production.
- The reliability bar for production trading is higher than most ML applications, and most ML systems do not clear it.
Where ML has earned a place in our stack
Machine learning has earned a defined, narrow role in our production stack. Feature extraction from order-flow data, calibration of conditional probability estimates, and regime classification are areas where ML genuinely outperforms classical statistics on our data. In each of these cases the role is to do something specific and measurable, with a held-out evaluation that has to be passed before the component goes live.
These are not headline-grabbing applications. They are infrastructure components — useful, narrow, and boring. That is exactly the profile we trust for production use. A component with a clear input, a clear output, and a measurable performance criterion is a component we can monitor and replace when it stops working.
Where ML has not earned a place
End-to-end deep models that take raw price data and emit positions have not survived our walk-forward tests. We have run honest experiments on architectures of varying sophistication. The in-sample plots are uniformly impressive. The out-of-sample behaviour, on data the model has never touched, is uniformly disappointing.
This is not a controversial claim if you talk to practitioners. It is, however, almost the opposite of the narrative sold publicly. The gap exists because in-sample backtests are cheap and impressive; honest walk-forward is expensive, slow, and frequently humiliating. The marketing surfaces the first; the production stack lives on the second.
“AI is a force multiplier for traders who already know what they are doing, and an expensive distraction for those who don't.”
Large language models in the research workflow
Large language models have a useful and growing role in our research workflow. Summarising filings, extracting structured data from analyst notes, drafting documentation, and generating first-pass code for new strategies are all areas where the productivity gain is real and measurable. The component does not need to be perfect; it needs to be a useful starting point for a human who will check the work.
What LLMs do not yet do in our book is generate signals in production. The reliability gap between 'useful research assistant' and 'component you trust with capital' is still wide enough that we do not bridge it. We expect this gap to narrow over time. We do not expect it to disappear in the near term.
The production reliability bar
Trading is a domain with an unusually high reliability bar. A model that is 95% accurate on a classification task is impressive in most applications and dangerous in a production trading system, depending on what the 5% failure mode costs. Most ML applications do not internalise this — the cost of a wrong prediction in a consumer product is annoyance; the cost of a wrong prediction in a trading system is capital.
This is why the components of ML that have earned a place in our stack are the ones with bounded downside on failure. A miscalibrated probability gets caught by the calibration monitoring layer and the component is paused. A misclassified regime gets caught by the supervision layer and exposure is reduced. End-to-end models do not have this kind of containable failure mode, which is the operational reason — alongside the empirical one — that we do not run them.
The pattern we expect to persist
The pattern we see across the industry is consistent. AI is a force multiplier for traders who already know what they are doing — the operational discipline, the risk framework, the execution layer — and an expensive distraction for those who don't. Throwing more model capacity at a trader without the operational base produces faster losses, not better trades.
We expect this pattern to persist even as model capability improves. The bottleneck is rarely the model. It is the operational system around the model, and that system is built with discipline, not with capability upgrades.
This piece is practitioner writing from a working self-trading desk. It is not investment advice. Defam AG trades only its own capital — see the disclosure page for the full statement.