A Hypothesis on Good Practices for AI-based Systems for Financial Time Series Forecasting: Towards Domain-Driven XAI Methods

Standard explainability methods have real limitations in finance: Popular XAI techniques like LIME and SHAP, while valuable, suffer from computational complexity, model bias, sensitivity to how data is sampled, and difficulty handling correlated features—issues that are particularly problematic in financial applications where accuracy and reliability are critical.

One-size-fits-all explanations don't work: The paper emphasizes that explainability methods must be tailored to specific audiences (regulators, customers, risk managers), suggesting that financial institutions need to develop audience-specific approaches rather than applying generic XAI tools uniformly across all stakeholders.

Data quality is foundational to trustworthy AI in finance: Good practices for responsible AI in financial forecasting begin with ensuring high-quality data, implying that even sophisticated explainability methods cannot compensate for poor underlying data—a critical consideration for institutions deploying these systems.

Explanation stability matters as much as accuracy: The paper highlights that explanations themselves must be stable and consistent, not just accurate. This means financial institutions should verify that their AI systems provide reliable explanations over time, preventing situations where the same input receives different explanations on different occasions.

Domain-specific XAI methods are necessary: Rather than relying solely on classical explainability techniques, the paper advocates for developing XAI approaches specifically designed for financial time series, acknowledging that finance has unique characteristics that generic methods may not adequately address.