WP-2026-001 · quant · ml
AlphaSeek - Regime-Aware Portfolio Allocation
Ibrahim Laklaa · 2026
0.242
best Sharpe (GMM-Markowitz)
69%
drawdown cut vs equal weight
13
market features
2009-2023
backtest window
Abstract
Detects latent market regimes with an HMM, a Gaussian mixture and spectral clustering on 13 engineered features, then conditions five allocation methods on the regime. Walk-forward backtests over 2009-2023: GMM-conditioned Markowitz doubles the Sharpe ratio of static Markowitz, and CVaR optimisation cuts the maximum drawdown by 69 % versus equal weight.
Question
Do markets move through distinct regimes, and does an allocation that knows the current regime beat one that does not?
Regimes as hidden states
Daily market features are modelled as emissions of a hidden state that follows a Markov chain:
Three detectors are compared on the same 13 engineered features (volatility, momentum, Sharpe, Hurst exponent, skewness and kurtosis, VIX, yields): a 4-state hidden Markov model, a Gaussian mixture and spectral clustering with an RBF kernel.
Allocation
For each detected regime, five methods are evaluated: Markowitz, CVaR minimisation, risk parity, Black-Litterman and equal weight. Mean-variance is the reference:
Protocol
Walk-forward backtests over 2009-2023 with rolling retraining: models only ever see the past, so there is no look-ahead bias.
Results
- GMM-conditioned Markowitz has the best Sharpe ratio, 0.242, against 0.091 for static Markowitz.
- CVaR optimisation has the lowest volatility (8.5 %) and cuts the maximum drawdown by 69 % versus equal weight.
- Every regime-conditional strategy beats the naive allocations.
Full tables and charts are in the 9-page paper.