Volatility · Advanced · Daily
VIX Mean Reversion
How it works, and when it doesn't.
The RF model's out-of-bag accuracy is 57% — applied with proper position sizing this edge compounds significantly. Feature importance shows RSI, 20-day return, and volume ratio are most predictive.
VIX=17.5 (z=-0.18) — no extreme fear reading. LR=69%. Waiting for spike signal.
Why it works
A 500-tree Random Forest trained on 50 technical features provides a data-driven probability estimate of the 5-day forward return direction. By aggregating thousands of feature interactions, it captures non-linear patterns that simple technical indicators miss. Walk-forward training prevents lookahead bias.
The signal logic
- 50 features: RSI, MACD, BB, volume ratios, multi-timeframe returns, VIX
- Walk-forward: model retrained monthly on 2-year rolling window
- Long: RF probability of up move > 60%
- Short: RF probability of down move > 60%
- Watch: max probability between 50-60% (uncertain)
When it fails
RF overfits to recent market regimes. Model trained in low-vol bull market underperforms during high-vol or bear regime.
- Best market
- Stable market regimes where feature-return relationships are consistent
- Worst market
- Structural breaks that differ dramatically from the training data
Backtest
| Sharpe ratio | 0.23 |
| Annual return | 1.9% |
| Worst drawdown | -14.6% |
| Trades in the backtest | 56 |
Backtest on daily bars, positions taken the bar after the signal (no look-ahead).
VIX Mean Reversion, 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.