Journal of African Development

In this research, we focus on analysing regime-dependent cross-asset correlations as a ploy that influences portfolio diversification from the perspective of normal market scenarios in congruence with crisis market conditions for the years 2010–2023. The gap created by the Modern Portfolio Theory (MPT) on the supposedly static correlation regime in the presence of empirical evidence suggesting structural breaks during market stress. We propose a new Markov-Switching Machine Learning (MS-ML) framework that integrates Dynamic Conditional Correlation (DCC-GARCH) models with XGBoost regime prediction and Deep Q-Network reinforcement learning for dynamic portfolio optimisation. A daily dataset of seven asset classes, including equities, bonds, commodities, real estate, and cryptocurrencies, is applied to rolling window analysis and ANOVA for regime comparison, along with true out-of-sample performance evaluations. The MS-ML framework shows an increase in risk-adjusted return to 45.6% from 80%, against the static mean-variance optimisation, marked by a 42.1% decrease in drawdown in crisis times. Consistent with the results, correlations worsen just in a crisis period, a mitigating factor that expels any diversification advantages exactly when they are of utmost importance to avoid riskier behaviour; our cryptos are not in safe-haven status. We, therefore, suggest that institutional investors should consider adopting a dynamic-regime-based allocation strategy with automatic rebalancing triggers. Similarly, risk managers should consider having dynamic hedge ratios of their own that automatically adjust before any days when volatility calls for action. Also, this study concludes that further investigations should dig deeper into daily intraday frequencies and incorporate ESG factors in regime-dependent optimisation