Predictive Policing
How statistical models and artificial intelligence use historical crime, location, incident, person, and environmental data to forecast where crime may occur or who may face elevated risk—and what agencies should understand about data quality, feedback loops, discriminatory effects, validation, transparency, explainability, officer discretion, civil rights, resource allocation, and the critical distinction between prediction and individualized legal suspicion.
What this explainer does
Predictive policing uses statistical analysis, machine learning, crime analysis, geographic information, or related analytical methods to forecast where crime may occur, identify people or situations believed to present elevated risk, or guide allocation of law-enforcement resources.
The concept ranges from relatively familiar crime forecasting— such as identifying locations where burglaries historically cluster— to substantially more controversial person-based systems that assign individuals risk scores or identify people believed more likely to become involved in violence.
The technology does not predict the future with certainty. It identifies statistical patterns in data and generates estimates about risk.
A model may conclude that a neighborhood faces elevated burglary risk or that a person shares characteristics associated with prior violence.
That statistical prediction is not itself proof that any particular person has committed, is committing, or is about to commit a crime.
1. Overview
Predictive policing applies historical and current information to statistical or computational models designed to estimate future crime-related risk.
The concept is not one technology. A simple statistical forecast of burglaries by block is fundamentally different from an algorithm that assigns individuals a violence-risk score.
Some systems are primarily resource-allocation tools. Others may influence surveillance, investigative attention, field contacts, intervention programs, or decisions concerning particular people.
2. Four Major Categories of Predictive Policing
| Category | Question | Example |
|---|---|---|
| Place-Based Prediction | Where is crime more likely to occur? | Forecasting burglary risk within small geographic areas during specified periods. |
| Person-Based Prediction | Who may be at elevated risk of offending or violence? | Risk scores based on criminal history, associations, victimization, or other variables. |
| Victimization Prediction | Who or what may be at elevated risk of becoming a victim? | Identifying locations or people at risk of repeat violence or burglary. |
| Crime / Offender Analysis | What patterns may connect offenses or help identify a perpetrator? | Using characteristics of crime scenes to identify likely relationships among incidents. |
3. How Predictive Policing Works
4. Place-Based Crime Forecasting
Place-based systems predict locations where particular types of crime may be more likely during a future period.
They build on longstanding criminological research showing that crime is not distributed evenly across geography and that relatively small locations can account for disproportionate numbers of reported incidents.
Identify geographic concentrations of prior crime.
Model whether a recent crime increases short-term risk nearby.
Analyze environmental features associated statistically with crime risk.
5. Person-Based Predictive Policing
Person-based systems attempt to identify individuals believed to face elevated risk of committing violence, becoming victims of violence, or otherwise becoming involved in specified criminal activity.
Inputs might include prior arrests, convictions, victimization, age, co-arrests, gang information, associations, geographic exposure, or other variables depending on the system.
A person can share characteristics associated with elevated risk without engaging in criminal activity.
Social or family relationships can become algorithmic variables even when the relationship itself is innocent.
Prior contacts or arrests may influence risk even when they did not result in conviction.
6. Predicting Victimization
Predictive analytics need not be used only to identify suspected offenders. Some systems attempt to identify people or places facing an elevated risk of victimization.
That information may support outreach, violence interruption, victim services, burglary prevention, directed patrol, or other preventive strategies.
7. Resource Allocation
Predictive systems can help agencies determine where limited patrol, investigative, or prevention resources might have the greatest effect.
That is one of the least controversial uses when the model operates at an aggregate geographic level and does not itself trigger coercive action.
Allocate officers to locations associated with elevated forecast risk.
Focus analysis on recurring locations or environmental factors.
Direct lighting, outreach, environmental design, or other preventive resources toward identified risks.
8. The Model Is Built from Data
Predictive systems do not discover crime independently. They learn patterns from the data they receive.
| Data Source | What It May Represent | Potential Limitation |
|---|---|---|
| Reported Crime | Incidents reported by victims, witnesses, or police | Not every crime is reported uniformly across communities |
| Arrests | People arrested for suspected offenses | Reflects enforcement decisions as well as underlying crime |
| Calls for Service | Requests for police assistance | Call patterns reflect reporting behavior and community-police relationships |
| Field Contacts | Police encounters and observations | Frequency can reflect prior deployment decisions |
| Geographic Data | Locations, businesses, transportation, land use | Correlations may not establish causation |
| Person-Level Records | Criminal history, associations, victimization, age, addresses | Errors or historical inequities can become model inputs |
9. “Dirty Data” and Historical Policing Patterns
Historical law-enforcement data are not neutral measurements of every crime that occurred in a jurisdiction.
They reflect what was reported, where officers were deployed, what conduct officers observed, which complaints were investigated, whom officers stopped, whom prosecutors charged, and countless other institutional and community factors.
10. The Feedback-Loop Problem
Predictive policing can create a self-reinforcing loop.
11. Correlation Is Not Causation
Predictive models identify statistical relationships.
A model may find that certain locations, histories, relationships, or other factors correlate with future events without proving that those variables cause crime.
12. What Does “Accurate” Mean?
Predictive-policing claims should be evaluated against defined metrics.
Of events predicted, how many actually occurred within the defined parameters?
Of events that occurred, how many were captured by the prediction?
How often does the system classify a location or person as high risk when the predicted event does not occur?
How often does crime occur outside the model's predicted risk?
Does a stated risk level correspond with observed outcomes?
Does using the prediction actually reduce crime, improve response, or produce another desired result?
13. Validation
Predictive systems should be tested before deployment and monitored after deployment.
14. Bias and Disparate Effects
Bias can enter predictive policing through multiple stages.
Prior policing patterns may shape the data used for training.
Available records may not represent all crime or all communities equally.
Variables may measure police activity rather than underlying criminal activity.
Statistical choices can distribute errors unevenly.
Police may respond differently to equivalent predictions in different areas.
Model-directed enforcement can create new data reinforcing existing predictions.
15. Proxy Variables
A model does not need to explicitly use race, religion, or another protected characteristic for those characteristics to be indirectly reflected.
Geography, housing, income, school, arrest history, associations, transportation patterns, and other variables can correlate with sensitive characteristics.
16. Explainability and Transparency
Agencies should understand enough about a predictive system to explain why its outputs are generated and how those outputs are intended to be used.
What information goes into the model?
What factors materially influence the prediction?
What exactly does a score or hotspot mean?
What causes a person or location to be categorized as elevated risk?
How frequently does the model change?
Under what circumstances should officers not rely on the output?
17. What Officers May Do with a Prediction
The constitutional and policy consequences depend heavily on how predictions affect officer conduct.
| Action | Prediction's Proper Role |
|---|---|
| Patrol Allocation | May help determine where officers spend discretionary patrol time. |
| Crime Prevention | May support environmental, outreach, or noncoercive interventions. |
| Investigative Prioritization | May help analysts decide where to look for additional evidence. |
| Investigative Stop | Prediction alone does not establish individualized reasonable suspicion. |
| Search | Model output does not independently create search authority. |
| Arrest | Prediction alone does not establish probable cause. |
| Use of Force | Force remains governed by circumstances confronting officers, not an algorithmic risk score. |
18. Fourth Amendment Considerations
Predictive analysis itself may involve information already lawfully held by government, but enforcement actions triggered by the prediction remain subject to ordinary Fourth Amendment standards.
Officers must still establish reasonable suspicion, probable cause, warrant authority, consent, exigency, or another recognized legal basis appropriate to the action taken.
19. First Amendment Concerns
Person-based forecasting can raise significant First Amendment issues if models use associations, social networks, protest activity, social-media behavior, group membership, religion, journalism, or political activity.
20. Due Process and the Right to Challenge Risk
Person-based systems can create difficult fairness questions when an individual does not know that government has assigned a risk score or placed the person on a high-risk list.
Errors may arise from mistaken identity, outdated criminal histories, incorrect gang information, stale addresses, inaccurate associations, duplicate records, or flawed data matching.
Does the person ever learn that the classification exists?
Is there a mechanism for correcting inaccurate information?
What government action occurs because of the score?
21. Transparency
Meaningful transparency does not necessarily require publication of every technical detail or security-sensitive feature.
But policymakers and communities should generally be able to understand what problem the system addresses, what category of data it uses, what type of prediction it makes, what officers may do with the output, and how the agency evaluates error and disparate effects.
22. Procurement and Vendor Questions
| Issue | Agency Question |
|---|---|
| Model | What statistical or machine-learning methodology generates predictions? |
| Training Data | What data were used to develop or calibrate the model? |
| Local Data | Which agency records become inputs? |
| Performance | What independent validation supports the vendor's claims? |
| Bias Testing | What geographic and demographic analyses have been conducted? |
| Explainability | Can the agency understand why particular outputs are generated? |
| Model Updates | Can the vendor materially alter the model without notice? |
| Audit | Are predictions, user actions, data changes, and model versions preserved? |
| Data Ownership | Who owns agency data, derived features, risk scores, and model outputs? |
| Termination | What happens to agency data and predictions when the contract ends? |
23. Agency Governance Framework
Specify whether the system predicts places, people, victimization, crime patterns, or resource needs.
Define what decisions predictions may and may not influence.
Document sources, quality, errors, retention, corrections, and legal authority.
Assess whether predictions or resulting enforcement burdens fall disproportionately on particular communities.
Determine whether police deployment is generating data that reinforces the model.
Test accuracy and operational effect rather than accepting vendor claims alone.
State expressly that predictive outputs do not replace reasonable suspicion or probable cause.
Require meaningful professional judgment before consequential use of algorithmic recommendations.
Apply heightened approval, review, correction, and audit procedures to individual risk scoring.
Explain the system's purpose, broad methodology, data sources, and safeguards where legally appropriate.
Preserve model versions, predictions, searches, user actions, and consequential decisions.
Periodically require affirmative review of whether demonstrated benefits justify continued use.
24. Questions Every Agency Should Answer
25. Where Predictive Policing Is Going
Models may increasingly update continuously using live CAD, ALPR, video, gunshot alerts, and other sensor information.
Predictive systems may become embedded directly into real-time operational platforms.
Models may combine text, video, audio, location, vehicles, networks, and historical records.
AI may move from predicting risk to recommending specific police responses.
Language models may summarize predicted risks and explain patterns to officers and commanders.
Agencies may increasingly combine crime records, intelligence, sensors, vehicles, social networks, and commercial data.
26. Key Terms
27. Related ShieldPST.ai Resources
Integrated data, sensors, cameras, analysts, and AI-assisted operational decision support.
Open explainer →Algorithmic identification, candidate lists, accuracy, human review, bias, and governance.
Open explainer →Public-source intelligence, online identity, automated analysis, and First Amendment considerations.
Open explainer →Vehicle-location data, automated alerts, historical searches, and investigative analytics.
Open resource →AI-assisted evidence analysis, lead development, human verification, and investigative safeguards.
Open resource →Return to the Shield Technology Reference Library.
Browse explainers →28. Selected Primary and Authoritative Sources
NIJ overview describing predictive policing as the use of advanced analysis and intervention models to forecast where or under what circumstances crime may occur and guide police resources.
Review NIJ overview
NIJ-sponsored research describing major predictive-policing categories, including place, offender, victim, and perpetrator prediction approaches.
Review research brief
NIJ discussion of the development of place-based crime analysis, forecasting, research limitations, and evidence concerning proactive policing.
Review NIJ article
DOJ report examining AI applications across criminal justice, including predictive policing, facial recognition, risk assessment, fairness, privacy, transparency, and responsible-use considerations.
Review DOJ report
NIST framework addressing AI governance, validity and reliability, transparency, accountability, privacy, security, and harmful bias.
Review NIST AI RMF
NIST guidance addressing how bias can arise from data, algorithms, human decision-making, societal structures, and deployment contexts.
Review NIST publication
NIST recommendations emphasizing the importance of testing, auditing, and field evaluation when law-enforcement agencies deploy AI systems producing predictive or analytical outputs.
Review NIST recommendations
Foundational research guide examining predictive methods, data, policing interventions, implementation, and evaluation.
Review research
29. Key Takeaways
- Predictive policing is not one technology. It includes place-based forecasting, person-based risk prediction, victimization prediction, resource allocation, and other analytical approaches.
- Predictive systems use historical or current data to estimate future risk; they do not literally know what will happen.
- Place-based forecasting generally presents different legal and governance risks from assigning risk to named individuals.
- Historical police data are not neutral measurements of all criminal activity. They reflect reporting patterns, police deployment, enforcement decisions, recordkeeping, and community behavior.
- Errors and historical inequities contained in source data can be reproduced or amplified by an algorithm.
- Predictive policing can create feedback loops when model-directed enforcement generates new police records that reinforce future predictions.
- Agencies should distinguish whether a model predicts crime or partly predicts where police have historically looked for crime.
- Correlation does not establish causation. Variables associated with crime risk do not necessarily explain why crime occurs.
- A model's accuracy should be measured through defined metrics such as false positives, false negatives, precision, recall, calibration, geographic performance, and demographic performance.
- Predictive accuracy does not by itself establish that deploying the system improves public safety.
- Agencies should independently validate predictive systems using conditions and data relevant to their actual jurisdiction.
- Removing race or another protected characteristic from the input does not automatically eliminate bias because other variables may function as proxies.
- Agencies should understand enough about a system's inputs, outputs, thresholds, methodology, and limitations to govern it responsibly.
- Most importantly, a predictive risk score does not constitute individualized reasonable suspicion or probable cause.
- Predictive information may help determine where officers deploy or where investigators look for evidence, but enforcement action still requires the independent legal basis applicable to that action.
- Person-based systems require heightened safeguards concerning First Amendment activity, inaccurate records, associations, notice, correction, retention, dissemination, and due process.
- Vendor contracts should address model methodology, data sources, validation, bias testing, model changes, auditability, data ownership, and termination.
- Agencies should evaluate the entire system: data → algorithm → prediction → officer decision → police activity → new data.
- The next generation of predictive policing will likely become deeply integrated with RTCCs, sensor networks, video analytics, ALPR, artificial intelligence, and automated decision-support systems.
- The central future governance question will therefore be not merely “Can the computer predict risk?” but “What may government do to a person because the computer says risk is elevated?”