We found that although human expertise can offer improvements beyond that available from a simple statistical model, these do not lead to systematic improvements. We proposed a fully automatable statistical promotional forecasting model that increased accuracy substantially, over baseline and expert forecasts. Further analysis showed that human expertise under conditions was still adding value to the forecasts, indicating a hybrid statistical-expert solution.

Sales forecasting is becoming increasingly complex due to many factors, such as shortening product life cycles, increasingly competitive markets and aggressive marketing. Often forecasts need to be manually adjusted, greatly complicating the demand planning process. Previous work has shown that up to 90% of forecasts may be adjusted in some industries, with promotions being the main reason. The overwhelming majority of adjustments are based on human expertise; however, there is little insight into the performance of such adjustments and what are the best practices.

The Lancaster Centre of Forecasting (LCF) conducted a review of the efficiency of judgementally adjusted promotional forecasts. We found that such adjustments can enhance baseline statistical forecasts, but not systematically, often failing in the situation where there seems to be the most evidence for a large increase in sales. Forecasters make unnecessary large positive adjustments. Our team developed statistical models attempting to improve on human judgement. We demonstrated that automatically specified statistical models for promotional forecasting can outperform on average human expertise and provide substantial improvements in terms of forecasting accuracy and robustness (around 30%).

Key features of the proposed model are its ability to handle multiple types of promotions and produce promotional forecasts even for products that have limited or no promotional history. Promotions often occur simultaneously, making it difficult to discern and model individual effects. We overcome this issue by transforming the input information in order to capture the salient features of the promotional mix. For products that do not have long promotional history, we infer archetypical effects by pooling information from established products and carrying this over to the new products. The result is accurate demand predictions.

Statistical promotional modelling can offer substantial improvements for the forecasting process, in terms of accuracy, robustness and resource efficiency. Based on this, the LCF team explored whether it is meaningful to replace human judgement with statistical promotional models. Our analysis indicated that although the overall improvements are in favour of the statistical approach, it is unable to capture fully human expertise, while a hybrid approach achieved the best performance.

The LCF is always keen to further its collaboration with industry. If you are interested for more information on this work please contact Dr Nikolaos Kourentzes.