All publications
Research · 2026
Adaptive Bayesian-ML Covariance Shrinkage (ABMCS)
A machine-learning-guided framework for dynamic estimation error control in MVO
Aries Harry Pratama · DeepMind Lab · ariesharry3@gmail.com
Portfolio optimisationMachine learningCovariance estimationIDXIndonesia
Summary
This paper introduces the ABMCS framework — a gradient-boosted ensemble that predicts the optimal covariance shrinkage intensity and target dynamically, conditional on detected market regime. Empirical validation on a 10-year, 20-asset synthetic multi-regime dataset demonstrates lower portfolio turnover (−4.1% vs Ledoit-Wolf), reduced maximum drawdown (−1.8%), and superior Sharpe ratio in Bull regimes (+0.135) with 81% regime detection accuracy.
Key results
Regime accuracy
81.0%
Turnover vs LW
−4.1%
Bull Sharpe Δ
+0.135
Max DD improvement
−1.8%
Mean α*
0.721
CV MSE
0.224
Out-of-sample comparison
| Method | Sharpe | Max DD | Turnover | Ann. vol |
|---|---|---|---|---|
| Equal weight | 1.135 | -10.10% | 0.0000 | 14.31% |
| Sample MVO | 1.11 | -13.37% | 0.1841 | 13.28% |
| Ledoit-Wolf | 1.128 | -13.09% | 0.1791 | 13.28% |
| OAS | 1.123 | -13.10% | 0.1818 | 13.28% |
| ABMCS (proposed)Proposed | 1.113 | -12.85% | 0.1718 | 13.22% |
Full paper
Open on ZenodoLoading paper…