How to Use AI Agents for Trading Strategy Research

A thirty-minute lightning lesson on using an AI research agent in trading strategy research: how to choose the next experiment from a result, what the agent can run for you, and what you still have to check before you trust its answer. A trading strategy is a chain of choices: the data and its timing, the labels, the features, the validation protocol, the model, the portfolio and the costs. The ML4T research workflow breaks into stages, each stage into activities and each activity into choices, about a hundred of them. A backtest reflects all of them at once, so a result is evidence, not a finding. The session works through one ETF result from the book: moving the label from a weekly to a monthly horizon scores better, and that is consistent with a real effect, with overlapping labels that overstate the evidence, and with one lucky fold. Each explanation predicts something different, so one experiment that changes a single choice against a fixed baseline separates them. A research agent is a language model with tools: it reads and writes project files, runs code and queries the registry of past experiments, and it reads short methodology files, called skills, when a method choice comes up. The book's saved ETF run shows one combining existing model predictions and comparing the ensemble with the baseline under the same backtest rules, end to end. What it cannot do is choose the experiment or certify its own result. The lesson closes with the failure modes to look for (a confabulated improvement, a run declared done too early, an inconclusive comparison reported as a finding), what to inspect (the specification, the data handling, the evaluation and the trace), and how the work divides between researcher and agent across the workflow. Questions at the end cover the outside-business-activity question for people working in finance, where the book covers research agents (Chapter 24), whether the code is Python only, and how to deal with agent hallucination. Stefan Jansen is the author of Machine Learning for Trading, whose companion code has passed 20,000 GitHub stars, and the founder of Applied AI, which builds production ML for investment teams. He has been a systematic investment partner since 2013 and has taught ML to more than 110,000 professionals through DataCamp and General Assembly. ------------------------------------------------------------ ML for Trading: From Research to Production - live cohort, Thursdays from December 3 https://maven.com/stefan-jansen/resea... ML for Trading in the Age of AI Agents - one-day live workshop, December 12 https://maven.com/stefan-jansen/ml4t-... ML for Trading: Foundations - self-paced https://maven.com/stefan-jansen/ml4t-... The book's code, including the saved research-agent run in Chapter 24: https://github.com/stefan-jansen/mach... CHAPTERS 00:00 How to Use AI Agents for Trading Strategy Research 00:20 Where this comes from 00:42 The book's code on GitHub 01:32 Choices, experiments and research agents 03:03 The ML4T research workflow 07:31 From idea to data, signal and strategy 09:13 Stages, activities and choices 11:52 Evidence, interpretation and decision 13:02 One ETF result, three explanations 15:24 Criteria for the next experiment 17:32 Research agents: tools and skills 19:39 An ML4T research agent at work 21:13 Checks before trusting an agent's result 22:37 Researcher and agent across the workflow 24:49 ML4T: Research to Production 26:40 Which route fits 27:39 Research to Production, cohort 3 27:56 Questions 28:29 Q&A: probabilistic versus deterministic methods 28:58 Q&A: is a course an outside business activity? 30:09 Q&A: which chapters cover research agents? 30:49 Q&A: is the code Python only? 31:24 Q&A: how to prevent agent hallucination 32:36 Closing