Evaluating Parallel Grid Search for Arima Hyperparameter Selection in Financial Time Series

Financial forecasting requires scalable time series modeling pipelines that balance predictive accuracy and computational cost. Although ARIMA remains widely used, exhaustive hyperparameter search becomes a major bottleneck in operational settings. This work proposes an automated ARIMA pipeline integrating stationarity testing with parallelized grid search for parameter selection. The study evaluates the computational and predictive impact of parallel hyperparameter optimization within a reproducible forecasting framework, using three Brazilian stocks from distinct market sectors (energy, utilities, and banking) spanning horizons from six months to eight years. Results show that parallelization achieves a geometric mean speedup of 1.74×, reaching up to 6.58× in medium-length series, while maintaining forecasting accuracy below 3% MAPE for medium-term horizons. Analysis based on Amdahl’s law indicates that parallel efficiency is strongly dependent on problem scale, with parallelizable fractions ranging from below 10% in very long series to over 70% in medium-length scenarios, where performance gains are most substantial. These findings demonstrate that parallel grid search can reduce computational costs in ARIMA deployment without compromising predictive performance, particularly when applied to datasets of sufficient scale.