Machine learning · Advanced · Weekly
Gradient Boosting Alpha
How it works, and when it doesn't.
When H > 0.55, trend-following strategies have historically shown 2x higher Sharpe ratios vs when H < 0.50. The Hurst Exponent is a leading indicator for which trading approach has edge right now.
GBM probability=53% — uncertain. No edge. Watch.
Why it works
The Hurst Exponent measures long-memory structure: H > 0.5 = trend-persistence (momentum works), H < 0.5 = mean-reversion (fade extremes), H = 0.5 = random walk (no edge). Rolling Hurst with Kalman smoothing provides real-time regime classification that selects the appropriate strategy type.
The signal logic
- Compute rolling 63-day Hurst Exponent using R/S analysis
- Kalman Filter smooths the noisy Hurst estimates
- Long when H > 0.55 (trending regime) AND price above SMA(50)
- Mean-reversion mode when H < 0.45
- Cash when H between 0.45-0.55 (random walk — no statistical edge)
When it fails
Hurst estimates are sensitive to window length. Short windows give noisy estimates; long windows are slow to detect regime changes.
- Best market
- Extended trending markets with persistent momentum (H remains high for months)
- Worst market
- Near-random-walk markets (H ≈ 0.5) where the signal oscillates without conviction
Backtest
| Sharpe ratio | 1.72 |
| Annual return | 20.6% |
| Worst drawdown | -7.3% |
| Trades in the backtest | 46 |
Backtest on daily bars, positions taken the bar after the signal (no look-ahead).
Gradient Boosting Alpha, 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.