Automatic model selection for time series forecasting: a sales data application

Demand forecasting in retail is increasingly relevant due to rising data availability and the growing expectation that artificial intelligence can make forecasting accessible to non-expert users. This thesis combines a theoretical foundation for time series analysis with a methodology that enables practitioners to generate reliable product-demand forecasts using only historical sales data extracted from invoicing systems. It addresses both the lack of analytical expertise and the heterogeneity of real-world time series. Throughout this work, the focus is on prioritizing simplicity, interpretability, and computational feasibility with the aim of enabling implementation in environments such as SQL. The proposed approach classifies time series by trend, seasonality, intermittency, and demand-size variability, and uses this structure to compare different forecasting models through a comprehensive walk-forward evaluation. The results identify the best-performing techniques for each class and show that no single technique is optimal across all categories: exponential smoothing performs competitively when a trend is present, Croston-type methods are effective for highly intermittent demand, and deep learning models achieve strong accuracy but require excessive computational resources. These findings support the design of an automated model selection pipeline that recommends appropriate forecasting models and provides interpretable reliability indicators for end users. The methodology, validated on sales data from a food and beverage company, offers a reproducible and operational framework that other organizations can adopt to improve forecasting efficiency. If provided by the company, future research could focus on integrating domain knowledge and incorporating exogenous factors to further enhance forecasting performance and decision-making usefulness.