Predictive policing software is a family of tools that sift past crime records to spot where crime tends to cluster, or who may be involved. Commanders use the output to place patrols ahead of events. The systems forecast patterns; they decide nothing about a person’s conduct.
Scale is what separates these tools from the paper crime maps departments kept for decades. Modern systems score city blocks, and sometimes people, from years of reports. Named independent reviews now ask two questions: do the forecasts work, and do they steer enforcement onto some neighborhoods more than others?
How does predictive policing software work?
The inputs are ordinary records: dates, places, and offense types from past incidents. The software hunts for clusters and repeats, then prints forecasts, usually as grids or lists ranked by risk. A department turns that output into a patrol plan, which typically sends officers to the estimated time and place of a crime.
Setup matters as much as the math: which data to collect, how often to refresh it, and how to judge whether an intervention worked. Those choices, not the model alone, decide what the system rewards in practice.
Analysts sort the tools into four groups, per a RAND Corporation review of the field:
- methods for predicting crimes, meaning places and times at raised risk;
- methods for predicting offenders, built on past records;
- methods for predicting who may have committed an unsolved crime;
- methods for predicting victims of crime.
What did the RAND review conclude?
The RAND review, published in September 2013, is a baseline assessment for the field. Its authors were Walter L. Perry, Brian McInnis, Carter C. Price, Susan Smith, and John S. Hollywood. Its key finding, per the report’s takeaways, is blunt: predictive policing methods are not a crystal ball.
They estimate risk from what came before. They cannot foresee what comes next. The review adds that the strongest programs sit inside larger strategies, ones that build ties between departments and their communities. Each method family also fits agencies of different sizes, from big offices with complex data to small ones with basic tools.
What did later audits and studies find?
The sharpest scrutiny concerns bias. A 2016 study by Kristian Lum and William Isaac, published in the journal Significance, ran a leading place-based algorithm on drug crime data from Oakland, California. The model sent police again and again to neighborhoods with more Black and low-income residents, while survey data showed drug use spread far more evenly across the city.
The authors traced the gap to bias in the arrest data feeding the model. Researchers call the mechanism a feedback loop. Past enforcement shapes forecasts, forecasts direct officers, and their new records feed the next round. Solon Barocas, Moritz Hardt, and Arvind Narayanan have argued that such loops are hard to fix with technical changes alone. This connects to our earlier piece, What police records management systems store, and why data quality varies.
Large datasets now let outside researchers test these questions at scale. The Stanford Open Policing Project has pooled more than 200 million records from dozens of state and local departments. The data cover vehicle and pedestrian stops, standardized so stop patterns and search rates can be compared across places rather than argued from single cases.
Do forecasts replace officer judgment?
They are built to inform it, not replace it. A forecast names places or time windows. Officers still decide where to go and what to do on arrival. RAND framed the systems as aids to proactive policing and to intervention plans, not substitutes for them. For related coverage, see How CAD systems route emergency calls, and what response time data misses.
That framing sets the audit standard. If the tool is an input, its errors can be measured. Biased inputs show up as skewed outputs, and accuracy claims can be checked against later records. The Oakland study and the pooled stop data are the working examples of that measurement.
Conclusion: Forecasts are evidence, not verdicts
Predictive policing software turns old incident records into risk rankings. The studies cut both ways: useful as one planning input, unreliable where the data carry bias, and never a substitute for a lawful decision by an officer. The audits that tested it asked the right question, which is whether the forecasts hold up outside the data that trained them.
