Is Big Data-Driven Predictive Policing Actually Effective?

Estimated reading time: 12 minutes

We are in a golden age of police technology.

Departments now have access to more information than ever before such as crime reports, calls for service, hospital data, census trends, even real-time environmental data. In theory, pulling these sources together should help agencies anticipate crime, allocate officers more precisely, and prevent harm before it happens.

That promise is powerful. If big data can truly sharpen predictions, it could mean fewer victims, smarter patrol strategies, and better use of limited resources.

At the same time, any technology that influences where officers go and how they engage communities deserves careful scrutiny. Big Data-Driven Predictive Policing (BDPP) has enormous potential, but it also raises real questions about accuracy, fairness, trust, and accountability.

To understand whether BDPP is living up to its promise, Youngsub Lee, Ben Bradford, and Krisztian Posch conducted a systematic review to examine how effective these approaches actually are when applied in policing practice.

What is Big Data-Driven Predictive Policing (BDPP)?

Traditional crime analysis usually looks backward looking back on yesterday’s calls, last month’s hot spots, or last year’s trends. BDPP is different. It combines multiple large datasets and uses algorithmic models to forecast where crime is likely to occur next, with the goal of shaping police activity before incidents happen.

BDPP is when police use many different sources of data and computer-based models to predict where crime is likely to occur, so resources can be deployed more strategically.

BDPP generally follows three steps:

1. Source – Police collect large amounts of data from multiple systems (crime reports, calls for service, demographics, hospital data, census data, and more).
2. Analysis – Algorithmic models analyze those datasets to identify patterns and generate crime forecasts.
3. Purpose – Agencies use those forecasts to guide patrol strategies and interventions.

This review focuses only on place-based BDPP, meaning the predictions are about locations rather than individuals.

The Study: The Evidence Base on Big Data-Driven Predictive Policing

To answer whether BDPP actually works, Youngsub Lee, Ben Bradford, and Krisztian Posch conducted a systematic review of the research literature.

They searched academic and policy studies published between 2007 and 2022 and identified 161 studies that examined big data–driven predictive policing.

Their research question was straightforward:

How effective is BDPP?

For this review, “effective” meant one of two things:

1. the approach led to measurable crime reduction when implemented operationally, or
2. the predictive model demonstrated credible accuracy in forecasting crime.

Out of the 161 studies, only six actually tested BDPP in real-world policing environments. The remaining studies were largely retrospective, evaluating whether models could predict crime rather than whether they produced measurable crime reduction.

What They Found

The authors sorted all 161 studies into four categories. But here is the key point:

Only six studies tested predictive policing in real police operations.
We will walk through those six first.

Before that, one key concept: displacement analysis.

Displacement analysis asks a simple question: Did crime actually go down, or did it just move down the street?

Without testing for displacement, an intervention may look successful even if it merely pushed crime into surrounding areas. Studies that include this step provide much stronger evidence that the strategy created a real net benefit.

That distinction is what separates Type 1 from Type 2 studies.

• Type 1 studies tested BDPP in practice and examined whether crime was displaced.
• Type 2 studies tested BDPP in practice but did not assess whether crime was displaced.

Type 1 – Real-world tests with displacement analysis

These are the strongest studies in the entire review.

Lowell, Massachusetts – Braga & Bond (2008)

Lowell PD blended 911 call data with officer insights about local conditions to identify 34 hot spots, covering just 2.7% of the city. Half were randomly selected for problem-oriented policing.

Over the next six months, calls for service dropped nearly 20% in the treatment areas. There was a slight increase in nearby areas, but it was not statistically significant, meaning crime did not meaningfully shift elsewhere.

Indianapolis, Indiana – Carter et al. (2021)

Instead of focusing on raw crime counts, Indianapolis targeted locations associated with the highest social harm meaning areas with the greatest monetary cost of crime and crime-related incidents. They identified these areas by combining police data, overdose reports, and crime cost estimates, and then compared them to traditional crime hot spots.

Those social harm hotspots experienced measurable reductions in social harm and violent calls for service. There was no evidence of displacement, and the strategy did not produce disproportionate arrest impacts on minority residents.

Philadelphia, Pennsylvania – Ratcliffe et al. (2021)

Philadelphia’s model combined crime data, demographics, and census income data. Officers tested three approaches: awareness only, marked patrols, and unmarked patrols.

Only one combination worked: marked patrols focused on property crime hot spots.

Property crime fell by 31%, with no displacement and even diffusion benefits to nearby areas. There was no effect on violent crime, and unmarked patrols produced no measurable change.

Type 2 – Real-world tests without displacement analysis

These studies implemented BDPP in practice, but without testing whether crime simply moved elsewhere.

Nashville, Tennessee – Wyatt & Alexander (2010)

Nashville used crash, DUI, and crime data to identify traffic hot spots. After implementation, fatalities and serious injuries declined, but there was no control group and no assessment of displacement.

Cardiff, UK – Florence et al. (2011)

Police and hospital data were used to identify violence hot spots. Serious injuries decreased significantly, although less serious injuries increased in the same areas.

Shreveport, Louisiana – Hunt et al. (2014)

Shreveport deployed predictive patrols across multiple target areas using several data sources such as historical property crime data, juvenile arrests records, disorder calls, and seasonal variation pattern information. The project failed to show meaningful results. The authors attributed this not to the model itself, but to organizational breakdowns and poor adherence to the intervention design.

Up to this point, the evidence looks cautiously promising. However….

The remaining 155 studies did not involve real-world interventions at all.

Type 3 – Retrospective accuracy studies

Type 3 studies made up the bulk of the research (134 in total), but they did not involve real-world police interventions. Instead, researchers looked at historical data to see whether new predictive models could forecast crime better than traditional methods. Essentially, they asked: If we had used this model in the past, how accurate would it have been?

This is important, but it is not proof that BDPP reduces crime. A model can perform well on historical data but still fail in practice. Factors like officer behavior, organizational constraints, community response, and fidelity to the intervention all influence whether predictions translate into real-world outcomes.

Still, these studies are useful because they show that building predictive models with large, multi-source datasets is technically feasible. Most suggest it is possible, but they also reveal that more data is not always better.

Key insights include:

• Combining multiple datasets often outperforms models based only on crime reports.
• How data is used matters more than how much data is used. For example, tracking who is present in an area at a given time rather than who lives there.
• Many novel variables, like weather, hotel reviews, taxi flow, or satellite imagery, often add little or no predictive value.

Takeaway: Type 3 studies don’t prove that BDPP reduces crime, but they show that sophisticated predictive models can be built and that careful selection of data matters more than simply adding volume.

Type 4 – Exploratory predictor studies

Type 4 studies were the smallest group (21 in total) and they did not build or test predictive policing systems. Instead, they looked at which factors tended to be linked with crime hot spots when combining multiple data sources. Their goal was to identify potential predictors that could later be tested in BDPP models.

Common categories of predictors included:

• Socio-demographic context: areas with high disadvantage, unemployment, housing instability, or frequent residential turnover.
• Built environment: locations with bars, liquor stores, convenience stores, transit hubs, mixed-use commercial areas, or clusters of vacant/abandoned properties.
• Signals of disorder and service demand: 311 requests, code-enforcement complaints, noise calls, and health indicators tied to assaults or substance use.
• Mobility and activity patterns: daytime population density, commuter flow, and short-term spikes linked to events.
• Policing and guardianship: patrol frequency, response times, and proximity to police facilities or cameras.

Type 4 studies offer the weakest form of evidence in the review. Type 4 studies are like scouting ingredients in the grocery store without ever cooking the meal. They tell us which items might matter (spices, produce, proteins), but they never show whether combining them actually produces something edible. Until those ingredients are tested in a real kitchen, we do not know whether the recipe works or fails.

What Can We Learn From Other Departments?

The Cautionary Tale: LAPD and PredPol

LAPD adopted PredPol with the same goal many agencies share: use advanced analytics to get ahead of crime and deploy officers more strategically. The system was designed to forecast short-term hot spots for property and violent crime, with the promise of helping the department focus limited patrol resources where they were most needed.

On paper, the potential was enormous.

In practice, the rollout revealed key risks of BDPP. PredPol relied heavily on historical crime data, which reflected decades of past policing. As a result, communities with higher levels of prior police contact were repeatedly flagged as future hot spots. This created a feedback loop: more patrols led to more recorded incidents, which then reinforced the model’s predictions.

Civil rights groups and journalists raised concerns that the technology was amplifying existing racial and geographic disparities rather than correcting them. Trust eroded, transparency was questioned, and the department ultimately discontinued the program.

The downfall was not simply the software. It was the combination of biased inputs, limited public accountability, and insufficient testing of how the model reshaped policing behavior.

A More Promising Example: Philadelphia

Philadelphia approached BDPP not as a replacement for officer judgment, but as a decision-support tool to be carefully tested. By pairing the predictive model with controlled, targeted interventions, the department found that marked patrols in property crime hot spots produced meaningful reductions, while other tactics did not.

The lesson: BDPP is not a silver bullet. When treated as an experimental tool with clear objectives and controlled testing, it has a much better chance of producing real results.

How to Test Big Data-Driven Predictive Policing in Your Agency

The takeaway from the research is clear: there is promise, but not enough evidence to know if BDPP or predictive policing will actually reduce crime in your context. Every agency is different; what worked in Philadelphia or Indianapolis may not work in yours. That’s why piloting and rigorous testing are essential before committing resources to full implementation.

If your department is considering BDPP or predictive policing, treat it as an experimental tool rather than a finished solution. A thoughtful pilot can help you determine whether the technology is effective, equitable, and operationally feasible.

Key elements of a strong pilot:

1. Use randomized control areas. Compare treatment areas with control areas to measure the real effect of the intervention.
2. Specify the crime type clearly. Target specific types of crime rather than general “hot spots” to increase clarity and effectiveness.
3. Measure displacement and diffusion. Track whether crime moves to nearby areas or whether benefits extend beyond the targeted locations.
4. Test officer compliance and fidelity. Ensure officers follow the intervention plan as designed; inconsistent implementation can obscure results.
5. Audit for bias. Examine whether the system disproportionately affects particular neighborhoods or demographic groups.
6. Be transparent with your community. Explain how data are used and how decisions are made to maintain public trust.

Strong recommendation: partner with an independent researcher or research institution. This helps ensure your pilot produces credible evidence, mitigates internal and external skepticism, and informs whether scaling up BDPP or predictive policing in your agency is truly worthwhile.

If you want a practical guide on how to design and execute a rigorous evaluation of a policing strategy from defining outcomes to measuring displacement and interpreting results. Read From Idea to Impact: Why Good Policing Strategies Fail and How to Make Them Work.

From Idea to Impact: Why Good Policing Strategies Fail and How to Make Them Work

Learn how Implementation Science helps agencies turn evidence-based initiatives into effective policing strategies that fit local realities.

Final Thoughts

BDPP holds real promise, but it is not yet fully proven.

Out of 161 studies, only six tested it in real-world policing. Five showed some effectiveness, while one failed due to implementation issues. That is not enough evidence to deploy the technology at scale without careful evaluation.

The real risk is not failure itself. It’s adopting BDPP or predictive policing as routine before understanding whether it actually works in your context.

The way forward is to test it thoughtfully within your agency. A well-designed pilot can show whether the technology improves outcomes, and whether it can be implemented fairly, effectively, and sustainably. That knowledge protects your resources, your officers, and your community trust.

Enjoyed this post?

Get more research translated into plain language delivered straight to your inbox twice a week.

Subscribe below and let me know which topics you want to hear more about. It only takes a minute.

If you found this helpful, share it with your team or on your socials. It helps more people make sense of research that matters.

Carlee Ruiz

Carlee Ruiz is the founder of PoliceResearchHub, a platform that turns academic research into practical insights for law enforcement. She brings seven years of experience as a crime analyst and justice researcher, with a focus on evidence based policing, gun violence prevention, and public safety strategy. Carlee currently works as a researcher at National Policing Institute, where she supports evaluations of law enforcement programs and criminal justice initiatives across the country. Through PoliceResearchHub, she aims to make research easier to understand and more useful for the people doing the work on the ground.