Supervised Learning (in Trading)

Supervised learning is the dominant machine learning paradigm in algorithmic trading because of its compatibility with the trading problem structure: 'given current conditions, predict future outcome'.

Problem formulation: given a feature matrix X (rows = time observations, columns = features) and target vector y (rows = time observations, values = future outcome), learn a function f(X) → y that minimises prediction error on out-of-sample data.

Common feature classes in trading ML: - Price-derived: recent returns over multiple horizons (1m, 5m, 30m), price relative to moving averages, distance from recent highs/lows - Volatility: realised volatility over recent windows, GARCH-modelled volatility forecasts, range-based measures - Volume: relative volume vs rolling average, volume-weighted price levels - Technical indicators: RSI, MACD, Bollinger Band position, stochastic oscillators (used as features, not as primary signals) - Time/regime features: time-of-day, day-of-week, days-since-last-news, regime classifications - Microstructure: bid-ask spread, order book imbalance (if accessible), recent trade direction

Common target variables: - Binary classification: 'up vs down' in next N bars - Multi-class: 'up significantly / unchanged / down significantly' - Regression: continuous return over next N bars - Triple barrier (López de Prado): outcome is whichever of [target_up, target_down, time_horizon] is hit first

Algorithm choices in order of typical use in retail/professional trading: 1. Gradient-boosted trees (XGBoost, LightGBM, CatBoost): the workhorse of tabular ML in trading. Handle non-linearities, interactions, mixed feature types well. Easy to interpret via feature importance. 2. Neural networks (MLPs, CNNs, LSTMs): used when feature engineering would be cumbersome (raw price sequences, chart images). Higher capacity but more overfitting risk. 3. Random Forests: less flashy than gradient boosting but more robust to overfitting on small datasets. 4. Logistic regression: simplest model; useful as a baseline. Often surprisingly competitive when features are well-engineered. 5. Support Vector Machines: less common today; out-performed by tree ensembles in most practical applications.

Validation challenges specific to trading: 1. Walk-forward validation (not random k-fold): time series cannot be shuffled. Must train on time period T, validate on T+1, retrain on T+1, etc. 2. Class imbalance: many trading targets have one class dominating (e.g. 'no significant move' is more common than 'big move'). Use stratified sampling, class weights, or threshold tuning. 3. Look-ahead bias: any feature computation must use only data available at the prediction time. Subtle bugs: using future volatility in z-score normalisation; using future high in target computation but training feature on similar future-derived values. 4. Non-stationarity: model performance often degrades as the market regime shifts away from training conditions. Periodic retraining is essential.

Integration with MT5 EAs: - ML inference can run in MQL5 directly for small models (decision trees, simple MLPs) - For larger models (neural networks, XGBoost), external Python service is more practical - Communication: MQL5 EA writes feature vectors to file or socket; Python service reads features, runs inference, writes prediction; EA reads prediction - Latency: file-based round-trip is typically 50-500ms — fine for hourly/daily strategies, problematic for scalping; socket-based is 1-10ms — usable for most strategies

For FxRobotEasy: our flagship AI EAs use supervised learning models trained on multi-year forex data with walk-forward validation. The models predict regime classifications (trend vs chop vs breakout) that inform parameter switching in the MQL5 execution layer.