Demand Forecasting of Individual Probability Density Functions with Machine Learning

TL;DR

This paper introduces a machine learning approach using Cyclic Boosting to predict complete demand probability density functions, enhancing explainability and reducing temporal confounding in retail demand forecasting.

Contribution

It presents a novel method for forecasting full demand distributions with explainability, addressing limitations of existing point estimate models and evaluation techniques.

Findings

Predicts individual demand distributions accurately

Provides fully explainable demand forecasts

Reduces impact of temporal confounding

Abstract

Demand forecasting is a central component of the replenishment process for retailers, as it provides crucial input for subsequent decision making like ordering processes. In contrast to point estimates, such as the conditional mean of the underlying probability distribution, or confidence intervals, forecasting complete probability density functions allows to investigate the impact on operational metrics, which are important to define the business strategy, over the full range of the expected demand. Whereas metrics evaluating point estimates are widely used, methods for assessing the accuracy of predicted distributions are rare, and this work proposes new techniques for both qualitative and quantitative evaluation methods. Using the supervised machine learning method "Cyclic Boosting", complete individual probability density functions can be predicted such that each prediction is fully explainable. This is of particular importance for practitioners, as it allows to avoid "black-box" models and understand the contributing factors for each individual prediction. Another crucial aspect in terms of both explainability and generalizability of demand forecasting methods is the limitation of the influence of temporal confounding, which is prevalent in most state of the art approaches.