Description:
Many forecasting methods perform poorly in the case of intermittent demand patterns and high variances in the demand quantity. Strong fluctuations within the time series and a large proportion of zero observations complicate the extraction of trend and seasonality. This research investigates how time series aggregation can improve forecasting models for time series with intermittent demand and high variations in sales quantity. We compare forecasting methods on a single time series with forecasting using time series aggregation as an additional regressor. We examine the differences between aggregation within predefined business groups from the business and aggregation using hierarchical agglomerative clustering on the sales history to overcome the lack of coherence within predefined product groups. Using a case study with more than 3,000 unique products of outdoor sports articles, we show that forecasting on a daily level can benefit from clustering depending on the time series characteristics. We show that the clustering approach outperforms the predefined product groups in almost all situations