Investigating explainable artificial intelligence in financial time series modeling

Artificial intelligence has become central to innovation in time series modeling, enabling increasingly complex and data-driven predictive models across a wide range of domains such as finance. This systematic literature review examines how explainable artificial intelligence (XAI) is currently applied in AI-based time series modeling, with a particular focus on financial applications. Following PRISMA guidelines, the review combines a comprehensive finance-specific search with a focused all fields search. It also combines descriptive analysis with qualitative synthesis to identify patterns in model adoption, the prevalence of XAI methods, and the limitations reported in the literature. The findings show that financial modeling largely adopts approaches originating from other application fields, with deep learning architectures dominating the landscape while established machine learning methods remain comparatively common in finance. At the explainability level, the literature is highly concentrated around a small set of post-hoc, model-agnostic techniques, most notably SHAP and LIME, which are frequently paired with dominant modeling architectures. However, commonly used implementations of these methods may rely on assumptions or approximation procedures that are difficult to reconcile with statistical properties frequently observed in financial time series, including temporal dependence and distributional change. The reviewed literature indicates that, under some implementations and data conditions, these mismatches may contribute to unstable, temporally misaligned, or potentially misleading explanations. By synthesizing these findings, the review highlights structural mismatches between generic XAI frameworks and the statistical properties of financial time series and outlines key directions for developing explainability methods better suited to sequential financial data.