RSight: A deep neural network for product demand forecasting over geographic regions

Hanjing Zhu, Sitan Yang, Tanmay Gupta, Harrison Waldon, Rahul Gopalsamy, Kin G. Olivares

Keywords: Time Series Forecasting, Deep Learning, Demand Forecasting

TL;DR: We propose a new DL model for product demand forecasting that shows superior performance over recent SOTA models such as Chronos and PatchTST.

Abstract: We propose RSight, a new deep neural network model for product demand forecasting across multiple geographic regions. Our model employs a novel region-enhanced encoder to learn cross-regional information. Using a dataset consisting of weekly sales for 15 million products from a large e-commerce company at the US Zip2 level, our method achieves substantial accuracy improvement over existing state-of-the-art forecasting models. Additionally, we demonstrate that RSight exhibits scaling effects with data sizes as we increase the number of series in our training population, we observe substantive performance improvements.

Submission Number: 19