Hierarchical Sales Forecasting In Multichannel Distribution Considering Marketing Campaigns

This paper focuses on demand forecasting for multichannel companies that rely on several distribution channels. Usually, multichannel distribution complicates forecasting because the demand patterns vary across the channels, especially due to the impact of marketing campaigns. In this paper, we take advantage of this circumstance. We interpret the sales data as a hierarchical time series, meaning we consider three different hierarchy levels: At the lowest hierarchy level, we have 20 time series that contain information on the sales of five different products per four different distribution channels. These time series can be aggregated to the middle level, which contain the sales figures per product. The single time series on the highest hierarchy level contains the sales of all products summed up.For every time series on every hierarchy level, we forecast the sales using both time series and machine learning models. Finally, we compare reconciliation methods that make the forecasts consistent among the hierarchy. We find that the reconciliation methods improve the overall forecasting accuracy.