<p>Accurate prediction of Wet Bulb Temperature (WBT) is critical for assessing heat stress and guiding early-warning and adaptation strategies in arid urban regions. This study explored advanced machine learning models for monthly WBT prediction in Jeddah City, Saudi Arabia, using long-term climatic observations spanning 1982–2023. Four models, namely, Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), Artificial Neural Network (ANN), and Random Forest (RF), were systematically assessed against both classical and climatological benchmarks. A seasonal-naïve baseline based on long-term monthly climatology explains 59% of WBT variability (RMSE = 2.38), while a univariate SARIMA model performs substantially worse (RMSE = 22.01; negative skill), highlighting the inadequacy of classical time-series approaches for capturing multivariate WBT dynamics. Among the individual models, RF and ANN markedly outperformed all baselines, with validation R² values of 0.995 and 0.997, respectively, whereas CNN exhibited consistently weaker performance. A year-wise bootstrap analysis reveals complementary strengths: RF demonstrates high temporal precision (RMSE = 0.148 ± 0.009), while ANN exhibited higher peak accuracy but greater interannual variability (RMSE = 0.064 ± 0.039). Motivated by this trade-off, we propose a hybrid ANN–RF model that combines ANN’s nonlinear representational capacity with RF’s stability, yielding improved generalization (validation RMSE = 0.150; R² = 0.997) and reductions of up to 28.6% in error relative to the best standalone models. Overall, the study’s findings demonstrate that hybrid machine learning models can deliver more accurate WBT predictions in arid regions, with direct implications for heat risk management, public health planning, and climate-resilient infrastructure design in rapidly warming arid cities.</p>