Previsão de séries temporais orientada por prompt com modelos de linguagem de grande escalas

Abstract

Time series forecasting is an essential tool for decision-making in various sectors, such as retail and urban mobility. This paper presents a comparative approach between traditional methods – represented by Random Forest and LSTM Neural Network (Long Short-Term Memory)– and an innovative methodology that uses Prompt Engineering-driven Large Language Model (LLM). For this study, two real datasets were used: one relating to product sales in a supermarket in Fortaleza and the other composed of boarding records on bus lines in the same city. While the classical methods employed the sliding window technique to structure historical data, the LLM-based method was developed through a prompt specifically designed for the Gemini-1.5-pro model, in order to capture seasonal patterns and trends in an integrated way. The evaluation of the models was carried out using the SMAPE (Symmetric Mean Absolute Percentage Error) and SEM Standard Error of the Mean), metrics, allowing a comparative performance analysis between the approaches.