Accurate long-term monthly temperature outlooks are vital for climate risk planning in the Gulf of Mexico and Atlantic-Transition. This study addresses the gap in comparisons between classical and machine learning methods by analyzing twenty-eight years of records from 1997 to 2025 across four coastal cities. Ten modeling families were benchmarked, including the following methods: Seasonal Autoregressive Integrated Moving Average (SARIMA), Holt–Winters (HW), Singular-Spectrum Analysis (SSA), Linear Regression (LR), Random-Forest Regression (RFR), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNN), and the hybrid Forecasting Method with Filters and Residual Analysis (FMFRA). The performance was validated using Symmetric Mean Absolute Percentage Error (sMAPE), Root Mean Square Error, and Directional Accuracy within a multidimensional ranking framework. Trend analysis revealed a statistically significant warming of 0.25 °C per decade. The FMFRA framework, integrating signal filtering and adaptive residual correction, achieved the best overall performance with an optimal Mean Rank of 1.75. Non-parametric statistical validation, conducted via the Wilcoxon signed-rank test with the Holm–Bonferroni step-down correction, confirmed that FMFRA consistently outperforms most machine learning and boosting architectures. While classical methods such as HW, SSA, and SARIMA remain competitive in stable maritime climates with low volatility, FMFRA provides superior robustness in regions characterized complex thermal transitions. Overall, integrating signal filtering with residual analysis yields more stable forecasts, offering a reliable computational foundation for proactive urban energy planning and climate risk mitigation in volatile coastal environments.
Frausto–Solís et al. (Wed,) studied this question.