How Faithful Is Attribution for Sales Forecasting? A Counterfactual Study

The researchers conducted a study on the posterior counterfactual interpretability of the multi-series WaveNet sales prediction model, training it using the Corporacion Favorita dataset (174,685 series, 1,688 days). The study evaluated attribution fidelity through deletion/insertion protocols and found statistically significant effects, proving that the attribution method reflects the actual behavior of the model rather than the allocation artifacts of traditional additive SHAP attribution. However, the analysis indicated that the attribution information was limited: promotional signal dependence varied across series (median ratio was approximately 1.0, with only about 20% of series showing significant effects). Additionally, although the model could capture the weekly sales cycle shape (daily-to-week correlation coefficient r=0.78), it systematically underestimated the amplitude of the sales cycle.