Abstract: Road icing is a physical process regulated by complex meteorological conditions and historical thermal inertia. The simplistic constructions of feature indicators and the class imbalanced samples are critical factors leading to missed detections in road icing forecasting. To enhance the forecasting performance of road icing prediction models, this study utilized meteorological observation data from Tianjin for the period 2020-2025. Drawing upon the concept of accumulated temperature, we proposed accumulated temperature feature indicators—such as the cumulative freezing index and effective freezing duration—to characterize historical thermal inertia. Consequently, a composite meteorological feature system for road icing was constructed, encompassing real-time meteorological features, process accumulation features, and physically derived features. By implementing random cross-validation based on precipitation processes, this study systematically evaluated the modeling performance of three machine learning algorithms: Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF), under two optimization strategies: cost-sensitive learning and composite feature integration. The underlying mechanisms of the optimal strategies for different algorithms were elucidated, and case verification was performed using the optimal model. The results indicate that road icing exhibits a lag effect of 0–12 hours following precipitation, and that process accumulation features and physically derived features can effectively distinguish between icing and non-icing processes. Composite features such as temperature–dew point spread, ground–air temperature difference, effective freezing duration, 12-hour accumulated precipitation, freezing accumulation index, and wind chill temperature demonstrate higher importance than traditional real-time features. Incorporating composite features based on prior knowledge can substantially improve the performance of LR and RF models. Furthermore, the adaptability of the optimization strategies is determined by the underlying mathematical logic of each algorithm. The LR and RF algorithm achieve optimal performance when integrating cost-sensitive learning with the composite feature strategy. In contrast, the SVM algorithm is susceptible to interference from loss weighting and high-dimensional features, resulting in performance degradation. All three models demonstrate good icing identification capability. Under their respective optimal training strategies, the models rank in descending order of overall performance as follows: RF, SVM, and LR. The optimal RF model achieves scores of ≥0.93 in precision, recall, F1-score, and Heidke Skill Score (HSS), enabling accurate prediction of road icing processes and providing scientific support for winter road traffic safety management.