Proposta de arquitetura de AutoML para aprendizado de múltiplos estimadores de séries temporais

Abstract

Demand forecasting in the retail sector is a highly complex task, characterized by the vast heterogeneity of patterns across thousands of time series. Traditional approaches, such as custom single models, are costly and poorly scalable, while global foundational models still face challenges in practical applicability. In this context, this thesis proposes and develops an AutoML methodology that enhances the robustness and computational efficiency of predictions, based on the cluster-then-forecast strategy. The cornerstone of this methodology is a novel clustering approach that employs SOM using the LikelihoodDistance metric to identify series with similar underlying generative processes. The architecture was validated in a real-world and challenging scenario, using sales data from the pharmaceutical retail sector. The results demonstrated that the sampling-based approach, derived from the clustering, was particularly effective, successfully identifying the best-performing models from small samples of series. The central finding of the research is that the proposed architecture not only proved to be competitive across all stores and evaluated metrics but also exhibited remarkable consistency and reliability. Furthermore, the success of the clustering provides strong evidence that the similarity between the series' generative models is a determining factor in selecting the most accurate forecasting technique. This work thus contributes an AutoML framework that mitigates the model selection problem in heterogeneous environments, offering a more stable, scalable, and computationally viable forecasting solution for the retail sector.