AI Signal Model Status

The EURUSD M15 LightGBM classifier is applied cross-symbol as an early preview. The feature set (RSI, MACD, ATR, BB, EMA crossovers, session flags) is computed per-symbol in real time, so inference runs on any instrument. However, the model was never validated on non-EURUSD data — signals outside EURUSD should be treated as experimental. Per-asset-class models trained and validated on each instrument family are scheduled for Q2–Q3 2026.

Minimum thresholds on the 180-day walk-forward test set: Sharpe ≥ 0.8, Win Rate ≥ 50%, Profit Factor ≥ 1.2, Max Drawdown ≤ −3R. Models that fail any criterion revert to staging for retraining. The metrics shown in the Model Registry are the values at the point of promotion — live performance is tracked separately in the Track Record.

Training uses OHLCV data for EURUSD M15 from 2015 to 6 months before the training start date. The last 6 months are held out as a walk-forward validation set. Triple-Barrier Method (López de Prado) labels are computed with a 1.5R reward:risk ratio and a time barrier of 48 bars.

The production model is retrained quarterly on an expanding window. The inference scheduler (PM2 ai_signals_inference) detects model version changes and hot-reloads the new booster without downtime. The MLflow run ID and version string in the registry are updated atomically on promotion.

An auto-rollback rule checks win_rate on a rolling 200-signal window. If it drops below 45% for 3 consecutive 24h windows the deployment is automatically reverted to the previous production version, a Telegram alert fires, and the new model reverts to staging for investigation.

Planned rows represent roadmap intent — the models have not been trained yet, so there are no acceptance metrics. Metrics will appear after training completes and the model passes acceptance criteria.

We use LightGBM's gain-based importance: the total information gain a feature contributes across all tree splits in the booster. A higher value means the feature had a stronger influence on the model's directional predictions.

Win rate alone is misleading. The system targets 1.5R or better on every trade. A 52% win rate with 1.5R per win and 1R per loss yields a positive expectancy. The Sharpe and profit factor metrics are more meaningful than raw win rate.

The model training pipeline is proprietary. The methodology is fully documented on the Methodology page and the feature set, labelling approach, confidence thresholds, and risk math are all described in enough detail to reproduce the core system independently.

Features are computed only from bars prior to the label window. The validation set (last 6 months) is never used during hyperparameter search. The triple-barrier labelling includes a time-based exit so future bar prices cannot leak into feature windows.

LightGBM on tabular financial data consistently outperforms deep models on moderate-sized datasets (< 5M bars). It trains in minutes, produces interpretable feature importances, and its DART mode handles the class imbalance of rare triple-barrier hits better than standard gradient boosting.

The three-class distribution (long / short / neutral) is roughly 20% / 20% / 60%. LightGBM's is_unbalance=True reweights classes automatically. Additional filtering by a confidence threshold (60% intraday, 55% daily) further reduces neutral noise from the signal feed.

M15 signals target the triple-barrier window of up to 48 bars (12 hours). In practice most trades resolve within 4–8 bars. A time-stop at bar 48 classifies the trade as expired (neutral) if neither TP nor SL was hit.