Interpretable Machine Learning for Promotional Sales Forecasting

Abstract

Background Sales promotions are marketing strategies used by retailers to stimulate the demand of products. They include a range of tactical techniques designed to encourage consumers to take direct and immediate purchase action. The impact of promotions in supermarket sales is substantial. The UK’s Competition Commission reported that sales promotions in British supermarkets typically lead to sales increases of 200%, with some offers reaching 3000%. Besides the increase in sales, promotions also attract customers into stores, increase store traffic, counteract competitors and reduce the retailer’s storage space and stockpiling. The estimation of promotional sales and its effect on other products is crucial, as it affects the whole supply chain. Objectives This Doctoral Thesis focuses on the study, development and validation of interpretable Machine Learning (ML) methods applied to promotions in grocery retail. In particular, it consists of the following three topics: (T1) promotional forecasting, (T2) cold-start promotional forecasting and (T3) quantification of sales cannibalisation due to promotions. The reason to focus on these topics specifically is that they are encountered by promotional forecasters on a daily basis. Also, that research in ML predictive methods has traditionally focused on error minimisation, whereas efforts in interpretability have not experienced the same surge. Finally, that interpretability of promotional predictions, as well as their effects, can benefit supply chain practitioners. The objectives common to the three topics are: (i) to develop and evaluate the methods on real-market data featuring different countries, store types and product categories, (ii) to evaluate interpretability on ground truth models, (iii) to benchmark our solutions against the ML state-of-the-art methods and to the retailer’s proprietary forecasts when available, (iv) to design our solutions to support forecasters in their decisions so that predictions can be easily modified, and finally, (v) to open source the implementation of all the methods in GitHub so they can be explored and used by the community. The objectives tailored to topic T1 are to discuss and evaluate the applicability of interpretable ML methods to the prediction of promotional sales. Along the same lines, topic T2 considers the problem of cold-start forecasting and its solution through interpretable ML. Topic T3 considers the problem of sales cannibalisation from a causal perspective.