Cluster-Based Forecasting for Intermittent and Non-intermittent Time Series

Tom van de Looij, Mozhdeh Ariannezhad

Abstract: Producing accurate forecasts is an essential part of successful inventory management for any retail business. Previous research has shown that the clustering of time series data into disjoint clusters can reduce the forecast error, eventually leading to cost savings. A common measure used to cluster time series data is Dynamic Time Warping. While it can handle time series of different length and guarantees to provide the optimal alignment, it is computationally expensive and assumes that one time series is a stretched non-linear version of another time series. For datasets containing intermittent time series, i.e. showing no clear structure, DTW is not the best suited method. In this paper, we propose a new framework that uses Simple Exponential Smoothing (SES) and a Self-Organizing Map (SOM) that is able to improve the clustering performance for clusters containing intermittent and non-intermittent time series. Using LightGBM as the forecasting model, we evaluate our approach on a real-world dataset, and find that the computational time can be reduced substantially compared to DTW when using a combination of SOM and LightGBM for both intermittent and non-intermittent time series while maintaining similar levels of accuracy.