Metropolitan crime in London (2010–2026): a reproducible LSOA-level analytics framework

This study analyses 16,969,327 Metropolitan Police street-level crime report records covering December 2010 to July 2026, obtained from the UK Police Open Data platform. After cleaning (removing records with missing spatial identifiers and duplicates), a final analytical dataset of 14,245,882 records across 16,110 Lower Super Output Areas (LSOAs) and 16 crime categories was produced. The research integrates exploratory data analysis, LSOA-level spatial aggregation, time-series visualisation, and interpretable machine learning using a Decision Tree Regressor. A temporal train-test split was employed, using records before July 2024 for training and the final twenty-four months for testing. This split was chosen, as opposed to a random one to avoid data leakage and as a factor of 12 months to address the seasonality of crime. Model performance was compared against a naive baseline (mean crime count per LSOA-crime type combination, computed from the training period). The Decision Tree achieved a Mean Absolute Error of 0.242 crimes per LSOA-month, outperforming the naive baseline’s MAE of 0.266, indicating that for long-term crime pattern predictions, tree-based regressors can provide a better performing tool for police resource allocation than simple averages. Feature importance analysis revealed that LSOA location plays the most important role, followed by temporal features. The crime type features were observed to have the lowest importance. The framework emphasises reproducibility, transparency, and operational applicability within a public-sector context.