Reinforcement learning (RL) trains an agent to take actions that maximize a reward over time. Instead of predicting a label, it learns a policy: in this state, take this action. That framing fits trading unusually well — and breaks in specific ways.

Where RL fits

RL shines on sequential decisions where the goal is clear: how to execute a large order with minimal slippage, when to scale out of a position, how to size given recent results. These are policy problems, exactly what RL is built for. NoVo, for example, uses a bandit-style approach to choose execution tactics — a close cousin of RL.

Why it's hard

Two problems. Reward design: if you reward raw profit, the agent learns to take insane risk for a lucky streak; you have to bake risk into the reward, which is subtle. Non-stationarity: markets change, so a policy learned on last year's regime can be actively wrong in this one (volatility regimes).

Reward the wrong thing and RL will optimize it ruthlessly — usually by taking risk you never intended. The agent does exactly what you asked, not what you meant.

The takeaway

RL is powerful for execution and sizing policies, less trustworthy as a standalone “what to trade” oracle. Like every AI tool, it belongs inside hard risk limits, not in charge of them — see AI signal vs AI execution and position sizing.