XCrime-LLM: An Explainable Spatio-Temporal Crime Prediction Framework

Crime prediction can support proactive public-safety planning, but practical deployment also requires outputs that are reliable and explainable. This study proposes XCrime-LLM for next-week crime occurrence prediction, in which engineered spatio-temporal features are serialized into a fixed prompt format and used to fine-tune GPT-4.1-mini to produce schema-guided JSON outputs from New York City Police Department (NYPD) incident records. The proposed XCrime-LLM framework is evaluated against prompting and trained baselines in New York City and further examined for cross-city transfer in Chicago. Supervised fine-tuning improved GPT-4.1-mini compared with the prompting baselines, increasing Micro-F1 from 0.7478 to 0.8095 and Macro-F1 from 0.7485 to 0.8075, while remaining competitive with the trained baselines. In the cross-city evaluation on Chicago, the fine-tuned GPT-4.1-mini outperformed the base GPT-4.1-mini without further fine-tuning or city-specific adaptation, raising Micro-F1 from 0.8277 to 0.8650 and Macro-F1 from 0.8693 to 0.9020. For explainability under black-box access, KernelSHAP identified last28_mean as the most influential feature across all crime types, while targeted ablation provided additional evidence of the model’s reliance on this feature. These findings suggest that the framework supports competitive next-week crime occurrence prediction while remaining explainable under black-box deployment constraints.