Reproduce “Demand Forecasting of Individual Probability Density Functions with Machine Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
F. Wick, U. Kerzel, M. Hahn, M. Wolf, T. Singhal, D. Stemmer, J. Ernst, M. Feindt
Researchers from Blue Yonder GmbH and IU Internationale Hochschule developed a method for forecasting individual customer demand probability density functions (PDFs) using a parametric Negative Binomial Distribution estimated by Cyclic Boosting. This approach explicitly mitigates temporal confounding and offers fully explainable predictions, achieving an Earth Mover Distance accuracy of 0.967 on retail sales benchmarks.
Key insights
- Predicting full probability density functions of demand allows for sophisticated optimization and strategic planning, going beyond the limitations of simple point estimates.
- Explicitly addressing temporal confounding leads to more robust, generalizable, and explainable models by disentangling true causal effects from spurious autocorrelation.
- The Negative Binomial Distribution is a more appropriate and robust parametric choice for discrete, over-dispersed retail demand compared to the Poisson distribution.
- Novel diagnostic evaluation techniques provide granular insights into model miscalibration, which is crucial for iterative refinement and building trust in probabilistic forecasts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.