Wielding Occam's razor: Fast and frugal retail forecasting

Abstract

Problem definition: Retail forecasting algorithms have increased in complexity in recent years, particularly those using machine learning. More traditional families of forecasting models, such as exponential smoothing and autoregressive integrated moving averages, have also expanded to contain multiple possible forms and forecasting profiles. Complexity in models, and in model availability, however, come at a cost and do not always offer accuracy or other benefits.

Methodology and results: Using a large scale empirical analysis, we show that one can parsimoniously identify suitable subsets of models without decreasing forecasting accuracy or a reduced ability to estimate forecast uncertainty. We propose a framework that balances forecasting performance with computational cost, resulting in the consideration of only a reduced set of models.

Managerial Implications: We demonstrate that considering a reduced set of models can perform well in retail practice and translate the computational benefits to monetary cost savings and improved environmental impact and offer implications to large retailers.

Keywords: exponential smoothing, ARIMA, big data, suboptimality, computational cost, retail, forecasting, forecast-value-added