Robert Fildes, Lancaster University, Lancaster, UK
Customer flow forecasting with GBRT: the benefits of adopting a customized machine learning approach.
With Shaohui Ma, Nanjing Audit University, Nanjing, 211815, China
Machine learning (ML) methods for time series forecasting have had a mixed (or worse) record of success compared to more standard statistical forecasting methods. This presentation will critically review the evidence before presenting a case study focused on forecasting many related series. The problem is concerned with forecasting customer flow, a key input for retailers when making their daily operational decisions. This is the first research utilizing the mobile payment data to provide participating stores with a value-added service by forecasting their daily customer flows, thereby overcoming the constraints small retailers face. We propose a third party mobile-payment platform centered customer flow forecasting solution based on an extension of the newly developed Gradient Boosting Regression Tree (GBRT) method which can generate multi-step forecasting for many stores concurrently. Using empirical forecasting experiments with thousands of time series, we show that GBRT together with a strategy for multi-period ahead forecasting provides more accurate forecasts compared with established benchmarks. Pooling data from the platform across stores leads to benefits compared to analyzing the data individually, demonstrating the value of this machine learning application. While hybrid ML methods have performed well in the recent M4 competition, this is the first research to offer convincing evidence over a large sample of time series of the potential strength of ML methods.
Robert Fildes is Distinguished Professor of Management Science in the Management School, Lancaster University, Director of the Lancaster Centre for Marketing Analytics and Forecasting and recipient of the Beale Medal, the UK Operational Research Society’s highest accolade. He was co-founder in 1981 of the Journal of Forecasting and in l985 of the International Journal of Forecasting. His current research interests are concerned with the comparative evaluation of different forecasting methods and the implementation of improved forecasting procedures and systems.