DAED: Dynamic Additive Effect Decomposition for Interpretable Time Series Forecasting

While predictive accuracy in time series forecasting is crucial, achieving model interpretability is equally essential in many high-stakes domains. However, most existing interpretable models focus on quantifying individual variable importance, overlooking the key role of variable interactions in capturing complex temporal dependencies. To address this limitation, we propose Dynamic Additive Effect Decomposition (DAED), a forecasting framework that explicitly models the importance of both individual and pairwise interactions in a dynamic and transparent manner. DAED decomposes each prediction into additive components of individual and interaction effects over time, providing a step-by-step interpretation of the forecasting process. Experiments on synthetic and real-world datasets demonstrate that DAED achieves competitive forecasting performance while quantifying the dynamic importance of individual variables and their pairwise interactions, offering a faithful and interpretable forecasting framework.