Machine learning · Advanced · Daily
Random Forest Classifier
How it works, and when it doesn't.
The MLP achieves 54% directional accuracy out-of-sample. By only trading when confidence > 60%, the effective accuracy on traded predictions rises to 58% — a meaningful edge compounded over many trades.
RF probability=50% — uncertain zone (0.45-0.55). No directional edge.
Why it works
A 3-layer MLP with walk-forward training on 30 technical features captures non-linear relationships for 5-day forward return prediction. Monthly retraining keeps the model adapted to current market conditions. Strict temporal separation prevents lookahead bias.
The signal logic
- Architecture: 30 inputs → 64 → 32 → 2 outputs (up/down probabilities)
- Features: returns, RSI, MACD, BB position, volume, VIX
- Walk-forward: train on 252d, validate on 63d, retrain monthly
- Long: P(up) > 0.60
- Short: P(down) > 0.60
- Watch: max probability between 0.50-0.60
When it fails
MLPs are prone to overfitting. Monthly retraining can be one cycle behind rapid structural breaks.
- Best market
- Stable market regimes with consistent feature-return relationships month over month
- Worst market
- Rapid regime changes between retraining periods
Backtest
| Sharpe ratio | 0.83 |
| Annual return | 8.8% |
| Worst drawdown | -9.1% |
| Trades in the backtest | 206 |
Backtest on daily bars, positions taken the bar after the signal (no look-ahead).
Random Forest Classifier, live.
Re-scored after every close.
The platform runs this and 32 other strategies against the live tape each day — with the evidence behind every number.