Robust Rolling Regime Detection (R2-RD)

A. Boukardagha & A. Saunders (Citi) • International Review of Financial Analysis

Abstract

R2-RD is an explainable framework for cross-asset portfolio optimization under time-varying market regimes. It combines expanding-window HMM estimation with: (1) dynamic BIC-based regime count selection, (2) a regime emergence policy ensuring monotonically non-decreasing counts, and (3) Hungarian algorithm label matching for temporal consistency.

Empirically (2016-2024): Sharpe 0.93 vs 0.73 for KNN, max drawdown -15.79% vs -30.44%.

Regime Detection Hidden Markov Models Portfolio Optimization Explainable AI

Performance (2016-2024)

Strategy	Return	Volatility	Sharpe	Max DD	Calmar
R2-RD + MVO	8.42%	9.05%	0.93	-15.79%	0.53
KNN + MVO	7.89%	10.81%	0.73	-30.44%	0.26
Equal Weight	5.12%	11.23%	0.46	-32.17%	0.16
60/40 Benchmark	6.78%	10.45%	0.65	-24.56%	0.28

Identified Regimes

Regime	Equity μ	Equity σ	Stock-Bond ρ	Interpretation
1	+1.2%	8.5%	-0.25	Low Volatility Bull
2	+0.8%	16.2%	-0.35	High Volatility Bull
3	-0.5%	18.7%	+0.15	Correction
4	-3.2%	32.4%	+0.45	Crisis

Methodology

HMM Observation Model: x_t | z_t=k ~ N(μ_k, Σ_k)
BIC Model Selection: K* = argmin { -2 log L + p · log T }
Regime Emergence: K_min,t = K*_t-1 (monotonically non-decreasing)
Label Matching: Hungarian algorithm for temporal consistency

Citation

@article{boukardagha2024r2rd,
  author    = {Boukardagha, Amine and Saunders, Alex},
  title     = {Explainable Regime-Aware Portfolio Optimization},
  journal   = {International Review of Financial Analysis},
  year      = {2024}
}