Machine learning · Advanced · Weekly
Ensemble Voting Classifier
How it works, and when it doesn't.
Static parameter optimization leads to gradual performance decay. Walk-forward adaptation maintains closer-to-in-sample performance because parameters are always tested on data they were never trained on.
Holding SPY
Ensemble weighted vote=0.59. LR meta=64%. Blended=70%. Sub-signals: [SMA=0(w=0.17), EMA=0(w=0.08), Momentum=0(w=0.08), WMA=0(w=0.08), DualMom=+1(w=0.59)].
Why it works
Walk-Forward Optimization continuously reoptimizes strategy parameters on rolling windows, then applies them out-of-sample. This prevents the gradual performance decay from static optimization as markets evolve. Online learning updates parameters weekly with new data.
The signal logic
- Rolling 252-day in-sample optimization window
- 63-day out-of-sample validation period
- Optimize SMA crossover parameters (fast: 10-30, slow: 30-100) on Sharpe ratio
- Apply optimal parameters to next 4 weeks of live trading
- Online learning: parameter updates weekly with new observations
When it fails
Over-optimization risk exists even with WFO if too many parameters are tuned simultaneously. Strict parameter bounds prevent worst cases.
- Best market
- Markets with slowly-evolving structure where 4-week-old parameters remain valid
- Worst market
- Markets that change faster than the reoptimization cycle
Backtest
| Sharpe ratio | 0.40 |
| Annual return | 4.6% |
| Worst drawdown | -25.5% |
| Trades in the backtest | 41 |
Backtest on daily bars, positions taken the bar after the signal (no look-ahead).
Ensemble Voting 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.