This study aimed to develop an intelligent and interpretable modeling framework for the monthly prediction of precipitation and temperature at 24 synoptic stations in northwest Iran over a 30 year period. For this purpose, the performance of three different approaches was evaluated: statistical modeling (Multivariate Linear Regression, MLR), fuzzy inference (Adaptive Neuro Fuzzy Inference System, ANFIS), and artificial intelligence (Multilayer Perceptron neural network, MLP). To overcome the limitations of classical methods in parameter optimization, a genetic algorithm (GA) was employed to design the optimal network architecture and determine the connection weights (resulting in the hybrid GA MLP model). The results showed that the proposed GA-MLP hybrid model delivered the most efficient performance by reducing the RMSE (e.g., demonstrating a 30% improvement in temperature and 23.5% in precipitation compared to the MLR model) and successfully overcoming the non-convergence issue of the baseline MLP model. Specifically, it lowered the temperature prediction error at the Parsabad station to 1.3°C and the precipitation prediction error at the Jolfa station to 10.9 mm. Sensitivity analysis using a feature importance approach revealed the key role of six month time lags and teleconnections in the region’s climatic fluctuations, confirming the model’s ability to grasp the underlying physics of the problem. Furthermore, uncertainty assessment using a calibrated bootstrap method indicated a high coverage rate above 95%, ensuring the model’s reliability for operational decision making. Finally, topographic analysis demonstrated no significant correlation between model error and elevation, which attests to the spatial robustness of the proposed model across heterogeneous climates.