Round 1: Model Development
Approach
Model: XGBoost for non-linear relation capture while maintaining efficiency
| Parameter |
Value |
| Tree Construction |
Histogram-based (faster, scalable) |
| Boosting Rounds |
300 |
| Max Depth |
6 |
| Learning Rate |
0.05 |
Features (14 total)
| Category |
Features |
| Momentum |
RSI, 7d Momentum |
| Moving Average |
MACD, EMA |
| Volatility |
Bollinger Bands, Volatility (1d, 7d) |
| Volume |
Average 7d Volume |
| Lagged Returns |
1hr, 4hr, 1d, 7d |
Results
| Metric |
Value |
| In-sample Weighted Spearman |
0.3113 |
| Out-of-sample Weighted Spearman |
0.1225 |
Gap suggests overfitting, but out-of-sample score remained strong given model simplicity.
Round 2: Trading Strategy
Design
Threshold-Based Approach:
- Trade only predicted returns > 2% magnitude
- 15-minute data (consistent with Round 1)
- Leverage model’s accuracy on extreme predictions
Position Sizing:
- Weighted by prediction magnitude (not evenly distributed)
- Example: +5% predicted BTC return → larger BTC allocation
Evolution
| Phase |
Issue |
Solution |
| Initial |
Underperformance after ~1 month |
Insufficient diversification |
| Fix |
Expanded trading list |
Include maximum perpetuals |
| Oct 10 Crash |
No extreme event handling |
Added stop-loss: -2% long, +2% short |
Future Improvements
- Replace fixed 2% threshold with volatility-based thresholds (Bollinger Bands)
- Implement ML-based stop-loss mechanisms
- Automate strategy parameters using indicators