Supervised-learning trading

Training a classifier or regressor on labelled historical bars to emit trading signals — the third answer to the decide layer, beside a learned policy (reinforcement-learning-trading) and a hand-coded rule set (the strategies banbot runs). Paged from Koshtenko’s XGBoost robot.

Where the labels come from — the whole problem

Supervised learning needs a target, and markets don’t ship one. The label has to be manufactured from the price series, and how you manufacture it silently defines the strategy. Koshtenko’s robot labels a bar by simulating the trade that would have followed it — 300-point stop, 800-point target — so “buy” means “a 300/800 bracket would have won.” Change the bracket and the labels change; the model learns a different game with the same features.

That’s the structural difference from RL. An RL agent gets a reward from interacting with a market and has to solve credit assignment across a whole episode; a supervised model gets a label someone defined per-bar and only has to fit it. Easier to train, and the difficulty moves into the labelling scheme, where it’s less visible.

Why the ensembles and the paranoid regularization

Financial features are weak, correlated and non-stationary, so the practitioner stack leans hard on variance reduction: gradient boosting under a bagging layer, recursive feature elimination down to a handful of inputs, and regularization strong enough to look like a typo ([[multithreaded-ml-trading-robot|subsample=0.01]]). The stated accuracy of 64–65% is worth reading against that: on a near-balanced binary label, the edge over a coin is small enough that the transaction-cost and slippage model decides whether it’s an edge at all — and neither is reported.

The standing hazard

Every supervised setup here trains on history and deploys into a distribution that moves. The corpus’s answers so far are all mitigations, not solutions: regime clustering to condition the model (market-regime-analysis), augmentation and synthetic data to widen the training distribution (synthetic-market-data), and honest out-of-sample discipline (backtesting). None of them makes the label definition correct.

Related

reinforcement-learning-trading · multithreaded-ml-trading-robot · backtesting · synthetic-market-data · market-regime-analysis · algorithmic-trading