A hierarchical agglomerative clustering for product sales forecasting

Many forecasting methods perform poorly when dealing with intermittent demand patterns and large variations in demand quantity. The presence of significant fluctuations within the time series, coupled with a substantial proportion of zero observations, complicates the extraction of trend and seasonality. This research aims to investigate the potential improvement in forecasting models for time series characterized by intermittent demand and high variations in sales quantity through the utilization of time series aggregation. Specifically, we compare forecasting methods on a single time series with forecasting using time series aggregation as an additional regressor. To address the lack of sales coherence within predefined business product groups, usually derived from the business context, we explore two distinct approaches to aggregation. Firstly, we examine aggregation within these predefined business groups. Secondly, we employ Hierarchical Agglomerative Clustering on the sales history to establish more coherent aggregation clusters. We conduct a comprehensive case study featuring over 3000 unique products within the outdoor sports articles domain. By examining forecasting at a daily level, we demonstrate that the effectiveness of clustering largely depends on the characteristics of the time series. Furthermore, we point out that the clustering approach consistently outperforms the predefined product groups in nearly all situations, thus affirming its superiority in enhancing forecasting accuracy.