Critical Evaluation of Time Series Foundation Models in Demand Forecasting

Abstract

Accurate forecasts are crucial as they enable organizations to make informed decisions about their supply chain. This research aims to benchmark and evaluate the efficiency of various foundation models in time series forecasting especially in the domain of demand forecasting. This research took two demand datasets from recent forecasting competitions and has used traditional statistical, machine learning and deep learning algorithms to forecast demand and compared their forecasting performance with popular foundational models TimeGPT and TimesFM. The evaluation considers both uncertainty and accuracy to establish a credible framework for comparison and benchmarking. This study has shown that TimesFM emerged as the better performing model across MASE & SMAPE and daily, weekly and monthly time granularities. The performance of the foundational models were at par with other traditional models and presented a strong case for wider research and adoption in industrial demand forecasting.