To address the insufficient integration of performance evaluation and contextual analysis in traditional architectural design, this paper proposes a design workflow that combines data-driven and performance-driven approaches, establishing a comprehensive operational pathway from typology selection and design generation to performance assessment. Using Yanshen Ancient Town, a cold region, as the study area, the research evaluates 18 traditional courtyard types and 8 brick kiln courtyard types. Benchmark models are selected based on the combined performance of PET (Physiological Equivalent Temperature) and MRT (Mean Radiant Temperature) indices. Subsequently, multiple performance indicators, including indoor and outdoor thermal comfort, indoor illuminance, and building energy consumption, are integrated into the analysis. Using a genetic algorithm, Pareto optimal solutions that meet performance requirements are iteratively optimized and filtered. Based on the learning rates and various evaluation indicators, XGBoost is ultimately selected to classify and predict the overall building performance. Results indicate that the model achieves an average prediction accuracy of 83.6%. Additionally, SHAP analysis of the independent variables in the algorithm reveals distinct influencing trends under different performance labels. The workflow demonstrates the feasibility of incorporating performance prediction in the early design stage of village courtyards, significantly enhancing the efficiency of feedback and follow-up between design decision-making and performance evaluation.