Traffic incident prediction is a key function of smart city management, allowing operators to proactively mitigate congestion and enhance roadway safety. While prior studies report promising results in single-city contexts, the generalizability of models across cities with distinct infrastructures, traffic dynamics, and data standards remains underexplored. This paper presents a cross-continental case study comparing Florida, USA (Traffic Message Channel data), and Tallinn, Estonia (DATEX II data). We introduce a harmonization framework to align structurally different datasets and evaluate three machine learning models: Extreme Gradient Boosting (XGBoost), Random Forest, and Long Short-Term Memory (LSTM) networks under both intra-city and cross-city settings. Intra-city experiments achieve accuracies of up to 94%, with XGBoost consistently outperforming alternatives. However, direct cross-city transfer without retraining results in a 13–15% F1-score decline, evidencing substantial domain shift. Feature-level analysis reveals that speed deviations and travel time anomalies generalize across contexts, whereas flow and temporal encodings are city-specific. The proposed harmonization and evaluation methodology establishes a reproducible basis for benchmarking across heterogeneous environments, underscoring the importance of local adaptation while guiding the development of transferable Intelligent Transportation System (ITS) solutions.