Machine learning maps fifteen years of London crime, street by street

London’s streets have generated an extraordinary archive of human misbehaviour. Between December 2010 and July 2026, the Metropolitan Police logged nearly seventeen million street-level crime reports, and a new open-access study in Computational Urban Science has now turned that sprawling administrative record into one of the most transparent crime-analytics pipelines yet published for any major city. The research, led by Rashid Nuraliev and Jas Semrl, analyses 16,969,327 records drawn from the UK Police Open Data platform and asks a deceptively simple question: can an interpretable machine-learning model, trained only on the past, predict how much crime each small neighbourhood will experience in the future better than a naive average can?

The answer, after a great deal of data cleaning, is a qualified yes. The raw dataset arrived as 188 monthly files spanning twelve structured variables, but it was far from analysis-ready. The authors first removed 2,565,351 exact duplicate records, roughly fifteen percent of the data, and then dropped a further 158,094 entries that lacked the spatial or temporal identifiers on which the whole analysis depends. That left 14,245,882 cleaned records covering 16,110 Lower Super Output Areas, the small statistical neighbourhoods used in UK census geography, and sixteen offence categories following Home Office classifications. Records were then aggregated into counts per LSOA, crime type and month, producing a spatio-temporal grid of more than fifty million potential combinations on which the models could be trained and tested.

The modelling design reflects a growing concern in applied machine learning: honesty about baselines. Rather than reaching for deep neural networks, the authors compared three deliberately explainable approaches. The first was a naive baseline that simply predicts, for each neighbourhood and crime type, the average monthly count observed during the training period. The second was linear regression, motivated by the apparently steady rise in reported crime since 2014. The third was a Decision Tree Regressor, chosen because tree-based models can capture non-linear relationships, such as the fact that crime hotspots cluster in central, eastern and western parts of London rather than varying smoothly along any single geographic axis, while still producing decision rules a human can read.

Crucially, the evaluation used a temporal rather than random train-test split. Records from December 2010 to July 2024, some 164 months, trained the models, while the final twenty-four months, August 2024 through July 2026, were held out entirely for testing. The authors argue this mirrors the real operational problem of forecasting future months from past data, and avoids the data leakage that random splitting would introduce when nearly identical records appear on both sides of the divide. The two-year test window is also an exact multiple of twelve months, so seasonal effects are balanced across training and evaluation.

On that held-out test set, the Decision Tree achieved a Mean Absolute Error of 0.242 crimes per LSOA-month, beating the naive baseline’s 0.266 and comfortably outperforming linear regression, which fared substantially worse than even the simple averages. The margins are small in absolute terms, and the authors are candid about that: at the granularity of a single neighbourhood and month, most locations record zero or near-zero incidents, so even modest improvements are meaningful. Error-distribution analysis showed the baseline systematically underestimating high-crime locations, whereas the decision tree, having learned spatial and temporal structure that simple averages ignore, did not. Both models, however, became similarly unpredictable at the most extreme one percent of error values.

Feature importance analysis from the tree revealed where its predictive power comes from. Spatial coordinates, the latitude and longitude of each neighbourhood’s population-weighted centroid, dominated at 70.9 percent of total importance. Temporal features, month and year, contributed 19.4 percent, while the one-hot encoded crime type features accounted for less than ten percent. Intriguingly, the x-coordinate proved a stronger predictor than the y-coordinate, suggesting that east-west variation across London influences reported crime levels more than north-south variation. Annual variation also outweighed seasonal variation, indicating that long-term trends, not recurring within-year cycles, explain more of the month-to-month differences the model must capture.

The descriptive findings are striking in their own right. Violence and sexual offences constitute the largest category, with 2,916,566 cleaned records, roughly one fifth of all incidents, followed by anti-social behaviour, other theft, vehicle crime and burglary. Spatially, the analysis confirms what criminologists call the law of crime concentration: a tiny fraction of micro-locations generates a disproportionate share of incidents. Westminster areas dominate the top-ranked LSOAs, with Westminster 018A recording 102,037 crimes over the study period, followed by Westminster 013B with 95,998 and Westminster 013E with 81,864. Seven of the ten most-affected neighbourhoods sit in Westminster, with the remainder in Hillingdon, Newham and Kingston upon Thames.

Temporally, the 188-month series shows a slight upward trend in incident counts after 2014, a dip in 2020 during pandemic lockdowns, and a rebound that, once linear growth is accounted for, does not fully return to pre-pandemic levels, a pattern the authors note is consistent with earlier research on COVID-era crime in London. Seasonality is unmistakable: averaging monthly counts across three years, July peaks at a mean of 81,200 crimes per month while February troughs at 69,390, an approximate seventeen percent seasonal swing, with summer months consistently recording peaks and winter months troughs.

The authors are explicit that their framework is a proof of concept with real limitations. The models exclude socio-economic covariates such as income, unemployment or population density; they count offences without modelling severity, victim demographics or clearance outcomes; and resident-based population normalisation was deliberately avoided because central London neighbourhoods swell enormously with commuters and tourists whose numbers resident counts cannot capture. Ensemble methods such as Random Forest and Gradient Boosting, which might improve accuracy further, were excluded on computational grounds and flagged as natural extensions, as are explicit time-series models, daytime and night-time population estimates, and empirical study of whether demand-prediction systems can create feedback loops that concentrate policing in already heavily patrolled areas.

That ethical caution runs through the paper’s discussion. Because crime records reflect historical patterns of reporting and policing activity, the authors warn that algorithmic predictions risk disproportionately directing resources toward already heavily policed neighbourhoods, and they argue that model outputs should support rather than replace professional judgment, with regular fairness evaluations. Within those boundaries, their conclusion is measured: among explainable approaches to forecasting policing demand, transparent tree-based models that beat both naive averages and linear regression offer public-sector analysts a defensible starting point, and the full pipeline, from open-data ingestion to temporal validation, is published so that other jurisdictions with comparable open crime data can replicate, scrutinise and improve upon it.

Subject of Research: Reproducible machine-learning analysis of sixteen years of London crime records at neighbourhood level

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

Article References: Nuraliev, R., & Semrl, J. (2026). Metropolitan crime in London (2010–2026): a reproducible LSOA-level analytics framework. Computational Urban Science, 6(1), Article 69. https://doi.org/10.1007/s43762-026-00305-1

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00305-1

Keywords: London, crime analytics, Metropolitan Police, machine learning, decision tree, LSOA, open data, spatial analysis, crime prediction, seasonality, Westminster, public sector