Impact of temporal granularity on machine learning models for time series forecasting

AbstractThis study examines the impact of timestep variation on the predictive performance of machine learning (ML) forecasting models, emphasizing the importance of optimal timestep selection for improved accuracy. The results show that Support Vector Regression (SVR) performs best with shorter timesteps but struggles with longer sequences. In contrast, Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks excel at 26 timesteps, leveraging their ability to capture patterns from extended contexts. RNNs demonstrate consistent performance across various timesteps, with their peak accuracy also observed at 26 timesteps. These findings highlight the need for careful timestep selection to enhance model efficiency and forecast reliability.