Volatility · Advanced · Daily
GARCH Vol Targeting
How it works, and when it doesn't.
LightGBM's leaf-wise tree growth captures interaction effects Random Forests miss. SHAP analysis shows VIX level, 20-day return, and RSI divergence are the top 3 predictive features.
GARCH(1,1) forecast vol=24.5% — elevated. Position reduced to 0.65x. Waiting for vol to normalize.
Why it works
LightGBM with Bayesian hyperparameter optimization learns complex non-linear relationships between 40+ features and forward returns. SHAP values provide full interpretability — you can see exactly which features drive each prediction.
The signal logic
- 40+ features: price, volume, volatility, macro (VIX, yield curve)
- Bayesian optimization tunes 8 hyperparameters
- Long: GBM probability > 0.58
- Short: GBM probability < 0.42
- Exit: probability reverts toward 0.5
When it fails
Without strict regularization and walk-forward validation, GBM overfits badly to recent patterns.
- Best market
- Markets with stable feature-return relationships (2019-2021)
- Worst market
- COVID flash crashes — unprecedented feature combinations the model has never seen
Backtest
| Sharpe ratio | 0.51 |
| Annual return | 5.6% |
| Worst drawdown | -20.6% |
| Trades in the backtest | 24 |
Backtest on daily bars, positions taken the bar after the signal (no look-ahead).
GARCH Vol Targeting, 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.