Deep generative models meet statistical methods: a generalized framework for financial regime identification

Regime identification in financial time series significantly impacts risk management and algorithmic trading but remains challenging due to the limitations of traditional statistical methods and deep generative models. To address it, we propose a generalized Mixture-VAE framework integrating structured probabilistic priors into the latent space of variational autoencoders (VAEs). The proposed approach simultaneously learns compact, informative representations and enforces temporal regularity through a rigorously derived state loss term. A key contribution of this work is the establishment of theoretical connections between the penalty terms in statistical jump models and the regularization in our Mixture-VAE, thereby providing novel interpretive insights into both paradigms. Our approach leverages the strengths of both statistical jump models and deep generative methods, effectively balancing interpretability with the expressive power of nonlinear embeddings. Empirical evaluations on synthetic and real-world intraday S&P 500 data demonstrate that our framework significantly outperforms existing methods, achieving superior performance in financial regime identification. Code is available at: https://github.com/yuqinie98/Mixture-VAE.