Trading Strategy
Most ML models applied to markets train one model on everything and hope it generalizes. This project challenges that assumption: first detect which market regime you’re in using a Hidden Markov Model, then route to a specialist Kolmogorov-Arnold Network trained only on days that looked structurally similar. KANs replace fixed activation functions with learnable spline curves, making the model’s logic directly readable. The result is a transparent, regime-aware system that surfaces cross-asset relationships invisible to single-model approaches.
The strategy underperforms buy-and-hold on raw return. The signal generator is conservative, spending ~56% of days in cash reduces volatility and max drawdown (−28.3% vs SPY’s −33.7%) but misses significant upside in sustained bull runs. This is a research proof-of-concept focused on interpretability and regime structure, not return maximization. The results are meaningful precisely because they are honest: walk-forward validation and careful leakage correction ensure no look-ahead bias inflates the numbers.
