A two-phased cluster-based approach towards ranked forecast-model selection

Abstract

Sales forecasting processes are usually automated to some extent in the retail sector and practitioners often have limited knowledge pertaining to the selection of appropriate forecasting methods. In this paper, we propose a generic two-phased, cluster-based framework capable of assisting retail forecasting practitioners in the selection of appropriate forecasting methods for time series of their retail sales data. One phase of the framework, called the benchmarking phase, involves establishing a benchmark data set (or updating it if it already exists) which can be leveraged to inform feature-based forecast model identification and ranking for different clusters of time series. The computationally efficient identification of a tailored shortlist of forecast models is thus facilitated during the other framework phase, called the implementation phase, for each sales time series presented to it by a retail organisation, based on the features of the time series presented. The two phases of the framework may be applied repeatedly in alternating fashion, enlarging the benchmark data set and improving its representativeness each time after having applied the implementation phase to the sales time series of a new retail organisation by re-applying the processes of the benchmarking phase. One iteration of this alternating application of the two framework phases is demonstrated and validated in respect of the M5 forecasting competition data (employed during the benchmarking phase) and a data set of the retail chain Corporacion Favorita (subsequently employed during the implementation phase).