New interpretability layer for sales forecasting models shows faithful attribution

Researchers have developed a new interpretability layer for deep sales forecasting models, specifically tested on a WaveNet-style convolutional network trained on grocery sales data. This post-hoc method provides counterfactual attributions that sum precisely to the predicted sales value, avoiding the artifacts seen in SHAP-style methods. The study rigorously evaluated the faithfulness of these attributions using deletion and insertion protocols, demonstrating a statistically significant effect that reflects genuine model behavior. The analysis also revealed that the model's reliance on promotion signals varies across different sales series and that while it captures the weekly sales cycle shape, it tends to underpredict its amplitude. AI

IMPACT Provides a more trustworthy method for understanding AI model predictions in sales forecasting, enabling better decision-making.

RANK_REASON Academic paper published on arXiv detailing a new method for model interpretability. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Favorita Corporation
• Google Wavenet
• Gotit.pub
• Hugging Face
• IArxiv
• ScienceCast

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →