Supervised vs reinforcement-learning models

Most AI trading systems sit in one of two camps: supervised learning or reinforcement learning. Knowing which you are looking at will sharpen every other question you ask.

A supervised model is trained on labelled examples. The data scientist hands the model thousands — sometimes millions — of historical windows along with the “right answer” for each (price went up, price went down, price did nothing). The model learns to map inputs to those labels. Once trained, it produces a probability for each label on new, unseen data.

Supervised models are easy to evaluate in the lab: you can compute accuracy, precision, recall and ROC curves on a holdout set. The trap is that price-direction labels are extremely noisy. A model can score 56% accuracy on a holdout and still lose money once spreads, slippage and position sizing are accounted for. Always ask whether the platform reports out-of-sample, fee-adjusted performance — not raw classifier metrics.

A reinforcement-learning agent learns differently. Instead of being shown the correct label, it is placed inside a simulated market and rewarded when its actions produce a desirable outcome (typically risk-adjusted profit). Over millions of simulated episodes, the agent develops a policy — a rule of thumb that maps any market state to an action.

RL has obvious appeal: it learns to act, not just to predict. The price you pay is enormous sample inefficiency and brittle behaviour outside the training distribution. An RL agent trained on a low-volatility regime can behave erratically when volatility spikes. Production systems usually constrain the agent inside a strict risk envelope (maximum position, maximum drawdown, kill switches).

When a platform markets an “AI bot,” find out which family it belongs to and what the training distribution covered. A supervised classifier trained on five years of bull-market data is a different beast from an RL agent that has seen 2008, 2020 and 2022 crises in its simulator.