Crime Prediction Modeling
Also known as: Predictive Policing, Crime Forecasting, Self-Exciting Point Process Crime Modeling, Predictive Crime Mapping
Crime prediction modeling forecasts where and when crime is most likely to occur next, so that limited resources can be directed before incidents happen rather than after. It spans simple historical hot-spot extrapolation, statistical self-exciting point processes that treat crimes as triggering further crimes, and modern machine-learning models that blend spatial, temporal, and environmental features. The statistical foundation was sharpened by Mohler and colleagues' 2011 demonstration that earthquake-style self-exciting (Hawkes) point processes — in which each crime raises the short-term risk of nearby crimes — forecast urban crime more accurately than conventional hot-spot maps.
Key highlights
- Forecasts the short-term space-time risk of crime, enabling proactive rather than reactive deployment.
- Self-exciting point processes capture near-repeat contagion explicitly and have outperformed conventional hot-spot maps.
- Flexible machine-learning variants can absorb rich spatial, temporal, and environmental features.
- Forecasts are evaluable on held-out future crime via PAI and PEI, allowing rigorous, comparable assessment.
- Integrates with hot-spot and risk-terrain analysis, combining where crime has been with where it is likely next.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use crime prediction modeling when you have a substantial history of geocoded, time-stamped crime and want to forecast near-term risk to guide proactive deployment. Self-exciting point-process models are well suited to crimes with strong near-repeat dynamics, such as burglary and gun violence; machine-learning models suit settings with rich features and abundant data. It is less appropriate when data are sparse or poorly geocoded, when the forecast horizon is long (most models predict days to weeks), or when historical data encode biased enforcement that the model would simply reproduce. Predictive systems also carry serious accountability and fairness obligations, so their use should be governed, transparent, and evaluated on held-out crime rather than in-sample fit.
Strengths & limitations
- Forecasts the short-term space-time risk of crime, enabling proactive rather than reactive deployment.
- Self-exciting point processes capture near-repeat contagion explicitly and have outperformed conventional hot-spot maps.
- Flexible machine-learning variants can absorb rich spatial, temporal, and environmental features.
- Forecasts are evaluable on held-out future crime via PAI and PEI, allowing rigorous, comparable assessment.
- Integrates with hot-spot and risk-terrain analysis, combining where crime has been with where it is likely next.
- Models learn from past recorded crime, so biased or under-reported enforcement data can be reproduced and amplified.
- Predictive accuracy decays with sparse data, poor geocoding, and longer forecast horizons.
- Strong in-sample fit does not guarantee operational value; only held-out hit-rate metrics reveal real performance.
- Self-exciting models assume a particular triggering structure that may not hold for all crime types.
- Deployment raises civil-liberties, transparency, and feedback-loop concerns that the statistics alone do not address.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What makes a self-exciting point process different from hot-spot mapping?
Hot-spot mapping summarizes where crime has concentrated and projects that forward, treating events as independent draws from a fixed density. A self-exciting (Hawkes) point process instead models each crime as raising the short-term risk of further crimes nearby, so the forecast updates dynamically after every incident, capturing near-repeat contagion. This dynamic structure is why Mohler and colleagues found self-exciting models forecast better than static hot-spot maps for crimes with strong triggering.
How is a crime forecast actually evaluated?
By how much future crime it captures in how little area, on genuinely held-out periods. The Prediction Accuracy Index (PAI) divides the share of crime captured by the share of area flagged, rewarding forecasts that pack a lot of crime into a small footprint, while the Prediction Efficiency Index (PEI) compares a model to the best possible capture given the same area. In-sample fit is not enough; only out-of-sample hit-rate metrics show whether a model is operationally useful.
What are the main risks of predictive policing?
The central risk is that models learn from recorded crime or arrest data that reflect where police have looked, not where crime truly is, so they can encode and amplify existing biases. Directing patrols by these forecasts can create feedback loops that generate more recorded crime in already-targeted areas, reinforcing the prediction. These concerns mean predictive systems require careful choice of inputs, transparency, governance, and evaluation of disparate impact, not just statistical accuracy.
Sources
- 1.Mohler, G. O., Short, M. B., Brantingham, P. J., Schoenberg, F. P., & Tita, G. E. (2011). Self-exciting point process modeling of crime. Journal of the American Statistical Association, 106(493), 100–108.
- 2.Perry, W. L., McInnis, B., Price, C. C., Smith, S. C., & Hollywood, J. S. (2013). Predictive Policing: The Role of Crime Forecasting in Law Enforcement Operations. RAND Corporation.ISBN 9780833081483
You have read it. What now?
Cite this page
ScholarGate. (2026, June 22). Crime Prediction Modeling. ScholarGate. https://scholargate.app/criminology/crime-prediction-modeling