Asian Journal of Probability and Statistics

Rapid urbanisation and land‑use change have intensified temperature variability in tropical cities, creating new challenges for climate resilience. Sri Lanka, which experiences heat stress across all districts, including Colombo (wet zone), Kurunegala (intermediate zone), and Puttalam (dry zone), requires reliable daily temperature forecasts to support agriculture, energy management, and public health planning. Using 4,916 daily meteorological observations (2010–2023), this study compares the performance of ARIMAF time‑series models with Multiple Linear Regression, Support Vector Regression (SVR), Random Forest, and Extreme Gradient Boosting (XGBoost). Among these approaches, SVR with the radial basis function kernel achieved the highest predictive accuracy (R² = 0.66 – 0.80; MAE = 0.34–0.44°C; RMSE = 0.45–0.57°C), while Random Forest and XGBoost achieved R² values between 0.60–0.75, outperforming ARIMAF (R² = 0.45–0.55), which showed weaker predictive capacity under nonlinear conditions. Feature importance analysis confirmed apparent temperature and reference evapotranspiration as dominant predictors across all districts, contributing >30% of the variance explained by the models. Unlike earlier Sri Lankan studies that emphasised seasonal or monthly analyses, this work provides the first systematic comparison of multiple machine learning and time‑series approaches for daily forecasting using high‑resolution daily data across all three districts. Although the limited availability of external datasets constrained external validation, methodological consistency was ensured through chronological train–test splitting and standardised evaluation metrics. The findings establish a robust framework for tropical urban temperature forecasting, offering practical insights for climate‑resilient urban planning and policy development.