Algoritmos de aprendizado online para previsão de séries temporais : uma análise do modelo ARIMA incremental

Abstract

This work investigated the performance of the incremental ARIMA model in online learning, addressing its adaptability and predictive power through simulations under various drift conditions and in the application to real data. Contrary to the initial hypothesis of sublinear growth, the results showed that both the mean and median regret grow linearly, yet with a coefficient significantly lower than α = 1, highlighting the model’s ability to maintain controlled regret growth and provide competitive forecasts compared to an ideally performing model. In the analysis using real data, the incremental ARIMA model exhibited performance similar to that of a traditional ARIMA model adjusted by moving windows, which performs an automated parameter selection. In situations of high volatility, both models faced challenges; however, the incremental version proved to be slightly superior in low to medium volatility contexts. Such results suggest that the incremental approach is capable of achieving performances comparable to methods that rely on complex parameter selection and adjustment, especially in dynamic environments that require quick adaptation.