Hierarchical time series forecasting via Support Vector Regression in the European Travel Retail Industry

Abstract

Times series often offers a natural disaggregation in a hierarchical structure. For example, product sales can come from different cities, districts, or states; or be grouped by categories and subcategories. This hierarchical structure can be useful for improving the forecast, and this strategy is known as hierarchical time series (HTS) analysis. In this work, a novel strategy for sales forecasting is proposed using Support Vector Regression (SVR) and hierarchical time series. We formalize three different hierarchical time series approaches: bottom-up SVR, top-down SVR, and middle-out SVR, and use them in a sales forecasting project for the Travel Retail Industry. Various hierarchical structures are proposed for the retail industry in order to achieve accurate product-level predictions. Experiments on these datasets demonstrate the virtues of SVR-based hierarchical time series in terms of predictive performance when compared with the traditional ARIMA and Holt-Winters approaches for this task.