Explainable Artificial Intelligence for Economic Time Series: A Comprehensive Review and a Systematic Taxonomy of Methods and Concepts

Machine learning models can predict economic trends better than traditional methods, but policymakers struggle to understand why they make specific predictions—this paper addresses that critical gap by organizing explanation techniques specifically designed for economic data.

Standard explanation methods often fail with economic time series because they ignore key features like autocorrelation and seasonal patterns; the paper identifies specialized adaptations (such as Vector SHAP and WindowSHAP) that preserve these temporal relationships while remaining interpretable.

For responsible policy use, the paper emphasizes connecting explainability to causal inference through methods like Causal Shapley values, enabling policymakers to distinguish between correlation and causation when using AI forecasts to inform economic decisions.

Attention-based transformer models offer a promising path forward because they are "intrinsically interpretable"—their internal mechanisms naturally reveal which economic factors the model is focusing on, reducing the need for add-on explanation tools that may introduce errors.

The paper provides practical guidance for "decision-grade applications," recognizing that economic AI systems used in real policy contexts require higher standards of transparency and reliability than research prototypes, with explanations that respect economic theory and constraints.