A machine learning framework for predicting weather impact on retail sales

Abstract

The weather affects the sales of many retail products worldwide. As the weather becomes more erratic due to climate change, retail organizations must respond by incorporating weather information into their sales forecasting models. This study proposes a modeling framework for identifying, quantifying, and evaluating the use of weather information in forecasting models. The models are developed using several time-shifted weather features and machine-learning techniques. Our method is applied to a dataset encompassing individual products and product categories obtained from a large Canadian retail organization. We find that using weather information improves the accuracy of sales forecasts significantly, explaining up to an additional 47% of the variance for the individual products and up to an additional 56% for the product categories, on top of the variance explained by a baseline model. By analyzing the parameters of the trained models, we can also determine the importance and influence of each weather feature, including time-shifted features. Our research findings contribute to both the literature on forecasting in the retail sector and the decision-making of retail organizations. By comparing a model developed with and without weather information, the organization can better determine the value of weather in its planning. Customer expectations of future weather significantly influence sales and should be considered for future studies. Our work provides a basis for researchers and retail organizations to forecast sales of individual products using weather information.