Track: Data Analytics

Abstract

The increasing variety in the retail sector and the complexity of the forecasting problems resulted in increasingly complex predictive models with difficult parameterization. When demand forecasts for a considerable number of items need to be performed, the forecasting process can take several hours or even days, which can be detrimental to business operations. This paper addresses this dilemma by presenting a model for forecasting retail demand in the health and wellness sector that improves forecast accuracy and balances computational cost and performance. The model incorporates classes of predictive methods using individual selection approaches and an aggregate selection where the method with the best performance for a population of time series is determined and applied. A time-series grouping technique was used due to the need to establish forecasts for more than one thousand SKUs, each one with demand characterized by an individual time series. Time-series centroids obtained by grouping with the DTW (Dynamic Time Warping) method were used to choose the best predictive method to be applied in each group. Results of this approach were compared with the help of the CHAID (Chi-square automatic interaction detection) algorithm. The selection of the predictive method through the cluster centroids showed a reduction in the average forecast error of 16.7% compared to the current model. In addition, the average processing cost of the proposed model was only 2.1 seconds per SKU, demonstrating that the selection of the predictive method for each cluster through the centroids proved to be an estimate with great application potential.