Abstract
Most systematic trading strategies apply fixed weights regardless of market conditions. A momentum fund stays long whether the market is trending or oscillating. This paper builds and tests a system that routes capital differently depending on which of four market regimes is active at any given time.
Two independent routing systems are built and compared on E-mini S&P 500 futures (ES): a deterministic paper router based on observable economic thresholds, and a machine-learning pipeline using an XGBoost gradient-boosted classifier trained on 74 engineered features. Both systems produce daily regime probability vectors that allocate capital across four specialist strategies covering trend-following, mean-reversion, defensive rotation, and crisis protection.
We develop a two-method soft-label framework for the ML system. Method 1 constructs labels from economically motivated feature scores passed through a softmax transform. Method 2 derives labels from forward-realised Sharpe ratios of the constituent strategies. Method 1 goes into production. Method 2 is dropped: bond-futures returns dominate its R3 (Choppy) labels, a structural contamination identified through SHAP analysis. We add three trend-persistence features after diagnosing an R1/R3 confusion problem in post-2021 low-volatility markets where ADX loses most of its discriminating power. These features were motivated by test-period observations; a fully clean evaluation requires data after May 2026.
On the out-of-sample test period (January 2021 to March 2026; 1,303 trading days for the paper router, 1,224 for the XGBoost production model), the volatility-targeted paper router achieves a Sharpe ratio of 1.47, CAGR of 12.30%, and maximum drawdown of 9.12%. The passive ES benchmark posts Sharpe 0.75 over the same period with a maximum drawdown of 25.02%. The XGBoost production model in soft-blending mode reaches Sharpe 0.81 with a maximum drawdown of 1.50%, its main contribution being capital preservation rather than return generation. Bootstrap confidence intervals place the paper router’s Sharpe estimate entirely above 0.8 at the 95% level; the White (2000) Reality Check p-value of 0.55 does not reach conventional significance, consistent with the router’s advantage arising from lower volatility rather than higher arithmetic mean returns. SHAP analysis computed on the full test set identifies ADX, HY OAS z-score, and the new trend_eff_20d efficiency ratio as the features most responsible for regime classification in out-of-sample conditions.
0 Comments