Regime Mamba: Regime Switch Detection in Financial Time Series via Mamba–Jump Hybrid Deep Model

Abstract

In this paper, we introduce Regime Mamba, a hybrid deep learning architecture combining the selective state space model (Mamba) with traditional Jump Models to identify financial market regimes. Although previous researchers on conventional regime identification have relied on traditional statistical approaches such as Hidden Markov Models or Jump Models, this is the first study to introduce modern deep learning methods to this domain. Through rigorous out-of-sample testing across multiple market conditions, our approach demonstrates superior performance, with 7.12% annualized return and 13.86% volatility compared to S&P 500 buy-and-hold strategies (18.82% volatility). Our model excels in risk management with a maximum drawdown of only -39.27% versus -56.83% for buy-and-hold, yielding a significantly improved Sharpe ratio of 0.4116. Thus, our approach shows exceptional cross-market generalizability, demonstrating effectiveness across both developed (S&P 500) and emerging markets (KOSPI).

Keywords

Regime Switching, Financial Time Series, Deep Learning, State Space Models, Jump Models, Risk Management