House-price prediction has been widely studied using structural and locational attributes such as bedrooms, bathrooms, floor area, and amenities. However, many existing studies underutilize environmental signals and rarely integrate long-term weather patterns within predictive frameworks. This study proposes an integrated machine-learning pipeline that combines textual housing attributes with granular climate variables to improve house-price prediction across six international markets, the United States, India, Vietnam, Saudi Arabia, Canada, and Pakistan. The methodology employs neural-network–based feature importance followed by iterative feature reduction and model training using Ordinary Least Squares, Random Forest, and XGBoost. Hyperparameter optimization is performed using Bayesian search with Optuna, and model interpretability is enhanced through LIME explanations. To further assess model robustness and environmental sensitivity, six diagnostic analyses are introduced: ablation analysis, causal sensitivity evaluation, SHAP-based interaction discovery, climate sensitivity metrics, stability analysis across random seeds, and uncertainty quantification using quantile prediction intervals. Results show that XGBoost provides the best predictive performance in five of the six countries, while Random Forest performs best in Saudi Arabia. Structural housing attributes remain the dominant predictors; however, climate variables significantly improve predictive performance and reveal region-specific valuation patterns. Weather effects are strongest in seasonal climates such as the United States and Canada, moderate in monsoon and tropical contexts such as Vietnam and Pakistan, and weaker in structurally driven markets such as India and Saudi Arabia. Overall, the findings demonstrate that integrating granular weather signals with housing attributes substantially enhances predictive accuracy and provides deeper insights into climate-sensitive real estate valuation.