Climate change is unarguably one of the biggest problems facing man in the twenty-first century. However, most recent research works on climate forecasting are based on traditional time series models which are unable to detect and forecast latent patterns inherent in climatic variables across global regions. In this study, we investigate the performance of autoregressive neural network algorithms using monthly temperature data of 30 years (1991–2020) obtained from the World Bank Climate Knowledge Portal. Clustered image generated by k-means clustering algorithm revealed the most similar climatic patterns among 32 countries in the Americas. Mann Kendall trend analysis revealed the presence of seasonal trend in the temperatures of all the countries considered except in Costa Rica and Bolivia. Autoregressive neural networks (ANNs) is then used to predict clustered monthly surface air temperature, and it was shown to outperform the traditional time series models of seasonal autoregressive integrated moving average (SARIMA) and Holt-Winters. Prediction accuracy of ANN differ among various regional temperature clusters with unequal sample sizes. Results from this study are useful for policy makers as well as climate modelers in selecting the best models for forecasting climate time series.