Reinforcement learning for trading

Reinforcement learning trains an agent to place trades by rewarding a goal such as profit or a risk-adjusted score. Academic papers have applied deep reinforcement learning to simulated markets and to historical price series. The published result is a research protocol, not a robot you can switch on. Costs, reward design, and the train-test split decide whether the agent learned anything.

Not investment advice. Nothing here is a recommendation to buy or sell anything. Backtests and historical studies are not guarantees, and a rule that worked in one sample can fail in the next. Trading can lose money. Links are for education, not endorsements. Do your own research.

Rules

Split history so the test period is untouched. Include commissions and slippage in the environment, not as an afterthought. Freeze the action space, such as long, flat, or short one unit, and the reward, before training. Report the test period once. Do not keep retraining until the test looks good, and do not deploy an agent whose only evaluation is in-sample equity.

Where AI fits

AI can implement the training loop, the cost model, and a clean test split, and it can compare the agent with a simple baseline such as time-series momentum on the same data. The useful role is the checklist from the papers: baseline, costs, and a true holdout. Asking a chatbot to 'use RL to make a strategy' skips that checklist.

Markets

Whatever liquid series the study specifies. A model trained on one futures market is not a model for another.

Timeframes

Whatever bar the study uses. Moving from daily training data to live ticks is a new problem.

Risks

Agents overfit, exploit simulator bugs, and collapse when the regime changes. Reward hacking can produce a policy that looks profitable and is not tradable. Deep RL is unstable across random seeds. Papers are not track records, and a hosted 'RL bot' selling access is not this research.

← Back to all strategies