ABSTRACT Understanding and quantifying the true effects of teleconnection patterns and the response time in which effects are observed on climate are major challenges for forecasting models. Thus, machine learning methods hold great potential to bridge this gap, supporting traditional climate models. Given the importance of climate prediction and its potential socioeconomic impacts, five different machine learning techniques were applied to forecast monthly precipitation and air temperature in different climate regions across Brazil, namely: Multiple Linear Regression (MLR), Elastic Net Regularisation (EN), Random Forest (RF), XGBoosting (XGB), and Support Vector Machine (SVM). The models were created using monthly teleconnection indices as well as time‐lagged temperature and precipitation measured at local surface meteorological stations. The data comprehends the period between 1979 and 2022, which were split between training (34 years) and testing (10 years) phases. Stationarity and diagnostic tests were also applied to identify potential fit failures in the models. Overall, all techniques exhibited good performance during the training phases, particularly the XGB and RF techniques, based on statistical indices. In the testing phase, the MLR models outperformed others in estimating temperature, while no consistent pattern emerged for the best technique in estimating precipitation across locations. This study demonstrates the potential of using diverse machine learning techniques to estimate meteorological variables in locations with varying climatic characteristics.
Freitas et al. (Mon,) studied this question.