Lifting the Veil of Non-Stationarity in Financial Market

Tianhao Fu, Xinxin Xu, Weichen Xu, Ruilong Ren, Bowen Deng, Xinyu Zhao, Jian Cao, Xixin Cao

Keywords: Non-Stationarity, Neural SDE, Jump-Diffusion, Financial Prediction

TL;DR: Combines explicit neural jump-diffusion SDEs with implicit multi-modal LLM reasoning to handle market non-stationarity and missing modalities for superior price movement prediction.

Abstract: Financial asset price movement prediction is inherently challenging due to the non-stationary nature of financial markets, where data distributions shift over time. Existing methods often assume that the market is stationary, which limits their applicability. To address this, we propose the Market-State Jump Diffusion Framework (MSJD), which models non-stationarity through two key components: an Explicit Market-State Jump Diffusion Process (EMJD) and an Implicit Market-State Jump Diffusion Process (IMJD). EMJD captures the dynamics of diffusion, drift, and jump processes governed by latent market states, formulated as stochastic differential equations, to explicitly model non-stationarity and solved via neural networks. IMJD integrates these components into a multi-modal large language model, enabling predictions across varying market conditions through temporal point encoding and jump diffusion embeddings to learn the non-stationary implicitly. Additionally, we introduce a general modality synthesizer that employs a unified adversarial masking strategy to complete missing modalities and fine-tune the prediction model. Furthermore, we validate our framework on the TSLA 2023 earnings crash, demonstrating that MSJD effectively identifies structural regime shifts. Extensive experiments on real-world stock and cryptocurrency datasets demonstrate that our method significantly outperforms existing approaches in the prediction of price movements. Project page and Code: https://iclr26-workshop-nonstationary.github.io/.

Submission Number: 118