Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Neural Additive Models for Interpretable Demand Forecasts

Leif Feddersen, Catherine Cleophas

TL;DR

Demand forecasting often forces a trade-off between accuracy and interpretability. We introduce Hierarchical Neural Additive Models (HNAM), which decompose forecasts into a level component and covariate effects governed by a user-defined interaction hierarchy, enabling additive, interpretable attributions while preserving predictive power. Empirical results on three retail datasets show HNAM delivers competitive accuracy relative to Temporal Fusion Transformers and consistently provides clear covariate attributions via its coefficient networks and level structure. The work highlights the practical value of inherently interpretable forecasts for human-in-the-loop decision-making and discusses challenges, such as causality interpretation and modeling delayed effects, alongside avenues for behavioral validation and future architectural enhancements.

Abstract

Demand forecasts are the crucial basis for numerous business decisions, ranging from inventory management to strategic facility planning. While machine learning (ML) approaches offer accuracy gains, their interpretability and acceptance are notoriously lacking. Addressing this dilemma, we introduce Hierarchical Neural Additive Models for time series (HNAM). HNAM expands upon Neural Additive Models (NAM) by introducing a time-series specific additive model with a level and interacting covariate components. Covariate interactions are only allowed according to a user-specified interaction hierarchy. For example, weekday effects may be estimated independently of other covariates, whereas a holiday effect may depend on the weekday and an additional promotion may depend on both former covariates that are lower in the interaction hierarchy. Thereby, HNAM yields an intuitive forecasting interface in which analysts can observe the contribution for each known covariate. We evaluate the proposed approach and benchmark its performance against other state-of-the-art machine learning and statistical models extensively on real-world retail data. The results reveal that HNAM offers competitive prediction performance whilst providing plausible explanations.

Hierarchical Neural Additive Models for Interpretable Demand Forecasts

TL;DR

Demand forecasting often forces a trade-off between accuracy and interpretability. We introduce Hierarchical Neural Additive Models (HNAM), which decompose forecasts into a level component and covariate effects governed by a user-defined interaction hierarchy, enabling additive, interpretable attributions while preserving predictive power. Empirical results on three retail datasets show HNAM delivers competitive accuracy relative to Temporal Fusion Transformers and consistently provides clear covariate attributions via its coefficient networks and level structure. The work highlights the practical value of inherently interpretable forecasts for human-in-the-loop decision-making and discusses challenges, such as causality interpretation and modeling delayed effects, alongside avenues for behavioral validation and future architectural enhancements.

Abstract

Demand forecasts are the crucial basis for numerous business decisions, ranging from inventory management to strategic facility planning. While machine learning (ML) approaches offer accuracy gains, their interpretability and acceptance are notoriously lacking. Addressing this dilemma, we introduce Hierarchical Neural Additive Models for time series (HNAM). HNAM expands upon Neural Additive Models (NAM) by introducing a time-series specific additive model with a level and interacting covariate components. Covariate interactions are only allowed according to a user-specified interaction hierarchy. For example, weekday effects may be estimated independently of other covariates, whereas a holiday effect may depend on the weekday and an additional promotion may depend on both former covariates that are lower in the interaction hierarchy. Thereby, HNAM yields an intuitive forecasting interface in which analysts can observe the contribution for each known covariate. We evaluate the proposed approach and benchmark its performance against other state-of-the-art machine learning and statistical models extensively on real-world retail data. The results reveal that HNAM offers competitive prediction performance whilst providing plausible explanations.
Paper Structure (57 sections, 4 equations, 3 figures, 5 tables)

This paper contains 57 sections, 4 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Motivation and Methodological Background
  3. Algorithm Aversion and Interpretability
  4. False Expectations
  5. Lack of Perceived Control
  6. Cognitive Compatibility
  7. Lack of Incentives
  8. Divergent Rationalities
  9. Judgmental Forecasting and Adjustments
  10. Methodological Background
  11. Prophet
  12. Neural Additive Models
  13. Neural Additive Time Series Models
  14. Ex-post interpretability
  15. Research Gap
  16. ...and 42 more sections

Figures (3)

  • Figure 1: Conceptual Overview of the HNAM architecture
  • Figure 2: Composed HNAM predictions and TFT predictions versus actuals in Walmart.
  • Figure 3: Composed HNAM predictions and TFT predictions versus actuals in Favorita.