Retail sales forecasting in the presence of promotions : comparison of statistical and machine learning forecasting methods

Abstract

Retail sales forecasting is the process of estimating the number of future sales for a specific product or products. However, producing reliable and accurate sales forecasts at a product level is a very challenging task in the retail context. Many factors can influence observed sales data at the product level, such as sales promotions, weather, holidays, and special events, all of which causes demand irregularities. Sales promotions are one of the salient drivers in generating irregular sales patterns. Sales promotions confound retail operations, causing sudden demand changes not just during the promotion period, but also throughout the demand series. As a result, three types of periods are relevant for sales promotions: normal, promotional, and a post-promotional. However, previous research has mostly focused on promotional and normal (i.e., non-promotional) periods, often neglecting the post-promotional period. To address this gap, we explore the performance of comprehensive methods, namely gradient-boosted regression trees, random forests, and deep learning in all periods. Moreover, we compare proposed approaches with conventional forecasting approaches in a retail setting. Our results demonstrate that machine learning methods can deal with demand fluctuations generated by retail promotions while enhancing forecast performance throughout all time periods. The base-lift model outperformed machine learning methods, although with more effort necessary to cleanse sales data. Our findings indicate that machine learning methods can automate the forecasting process and provide significant performance even with the standard approach. Hence, our research demonstrates the way retailers can successfully apply machine learning methods in forecasting sales.