Análise de estimativas automatizadas do preço de ativos financeiros utilizando ARIMA com grid search paralelo

Abstract

Time series forecasting tools are an essential component of the financial market, as they enable investors to make informed decisions based on historical data and enhance their profits. Among the traditional models that underpin these tools, ARIMA stands out for its ability to produce accurate forecasts at a relatively low computational cost. In this work, an automated ARIMAbased forecasting pipeline is proposed, aiming to reduce manual intervention in the selection of model parameters. To this end, the Augmented Dickey–Fuller test is applied in conjunction with a parallel grid search, which automatically determines the model order (p,d,q) to be trained. Model training and forecasting are performed using a rolling forecasting strategy. To validate the proposed method, experiments were conducted using data from Petróleo Brasileiro S.A. (Petrobras), Transmissora Aliança de Energia Elétrica S.A., and Itaú Unibanco, comparing the performance of the parallel approach with its sequential counterpart. Forecasting accuracy was evaluated using MAE, MAPE, and RMSE, along with computational performance, achieving an average speedup of 1.92× during the grid search. Although limitations of the ARIMA model and parallelization for long time series (eight and three years) were identified, its performance for medium-sized series was satisfactory, achieving good predictive accuracy.