Regime

We built a quant trading system for the SG × HK university Web3 quant hackathon that ranks all 66 tradeable coins every two hours and rotates capital into the strongest, using a regime-aware ML model and three tiers of risk control. In a walk-forward out-of-sample backtest from Sep 2025 to Mar 2026, BTC fell 37.9% while the strategy returned +14.7% with an 8.1% max drawdown.

Highlights

• Regime detection with a 3-state Gaussian HMM, feeding per-regime LightGBM rankers over 45 features — technicals, cross-asset BTC signals, funding rates and the Fear & Greed index
• Trained and shipped the ML ensemble myself (AUC 0.72 in trending regimes, 0.67 ranging) — it had never run in the live bot before
• Found that the circuit breaker and daily loss limit were never wired into the buy path, and aligned live stops to the thresholds the backtests had actually tested
• Dynamic sizing from equity-curve scaling, conviction weighting and pyramiding into winners, with rotation hysteresis that cut trade count ~35%
• Three-tier risk engine: hard and trailing stops, flash-crash exits, a 12% drawdown circuit breaker and a 5% daily loss limit