Data-driven forecasting of sales influenced by climate variability using deep learning

Abstract

This study investigates the statistical and machine learning models to improve sales forecasting accuracy in the retail sector by incorporating both transactional and environmental variables. Motivated by the limitations of traditional data handling systems within a case study company, the research proposes an integrated, centralized information system that enhances data accessibility, reduces redundancy, and supports timely decision-making. Multiple forecasting approaches—including SARIMA, SARIMAX, LSTM, Ordinary Least Squares (OLS), and Poisson regression—were evaluated using historical sales and weather data. Results indicate that regression-based models (OLS and Poisson) outperformed time series and deep learning models in terms of model fit and predictive power, emphasizing the effectiveness of simpler, interpretable methods when relevant features are included. The study also demonstrates that weather conditions, such as humidity and temperature, exhibit moderate correlations with sales volume, though their direct predictive contribution is limited when used in isolation. This data-driven framework offers a scalable solution for retail operations, contributing to cost reduction—such as minimizing reliance on third-party business intelligence tools—and promoting sustainable competitive advantage.