Predictive Policing in America - Criminal Justice

Predictive policing in America refers to the family of algorithmic and statistical forecasting tools American police departments have adopted since roughly 2010 to anticipate where and, in some implementations, by whom future crime is likely to occur, extending the place-based crime concentration research examined throughout the Environmental Criminology silo into forward-looking operational forecasting rather than purely retrospective analysis. This article traces predictive policing’s theoretical foundations in crime concentration research, the major algorithmic approaches departments have adopted, the empirical evidence evaluating these tools’ forecasting accuracy and crime-reduction effectiveness, the substantial equity and civil liberties controversies that have accompanied predictive policing’s expansion, and the resulting wave of department-level program cancellations that has reshaped the field since the mid-2010s.

Introduction

Predictive policing emerged directly from the empirical foundation the crime concentration research examined in the companion article on Crime Concentration elsewhere in this silo established, since David Weisburd’s documented finding that high-crime locations display substantial year-over-year stability supplied the essential statistical precondition for forecasting: a genuinely predictable, rather than randomly fluctuating, underlying pattern (Weisburd, 2015). Where the hot spots policing examined in the companion article on Hot Spots Policing elsewhere in this silo allocates resources based on documented historical concentration, predictive policing extends this logic by applying statistical and machine-learning models to forecast where concentration is likely to shift or intensify in the near future, a forward-looking analytical ambition that has generated both considerable practitioner enthusiasm and sustained civil-liberties controversy since the technology’s initial widespread adoption.

This article examines predictive policing’s theoretical foundations, the major algorithmic approaches departments have implemented, the accumulated empirical evidence regarding forecasting accuracy and crime-reduction effectiveness, the equity and civil liberties concerns that have generated sustained controversy, and the significant wave of program cancellations that has reshaped the field over the past decade.

Theoretical Foundations

Extending Crime Concentration Research Into Forecasting

Predictive policing rests theoretically on the same routine activity and place-management mechanisms examined in the companion article on Place and Crime elsewhere in this silo, extended through the additional empirical claim that these mechanisms operate with sufficient temporal consistency to support genuine forward-looking prediction rather than merely retrospective description (Cohen & Felson, 1979). Weisburd’s proposed law of crime concentration at place supplied particularly direct theoretical support for this predictive extension, since a jurisdiction-specific concentration pattern proving stable across multiple years, as Weisburd’s Seattle research documented, implies that near-term future crime distribution should be forecastable with reasonable accuracy from historical spatial patterns alone (Weisburd, 2015).

Elizabeth Groff, Weisburd, and Sue-Ming Yang’s longitudinal trajectory research supplied additional theoretical nuance relevant to predictive policing’s foundations, documenting that while most street segments displayed stable long-term crime trajectories, a meaningful minority displayed more volatile, less predictable patterns, implying that predictive accuracy should be expected to vary considerably across different locations within the same jurisdiction rather than achieving uniform forecasting reliability throughout (Groff, Weisburd, & Yang, 2010).

Near-Repeat Victimization Theory

Beyond general crime concentration stability, predictive policing algorithms frequently incorporate near-repeat victimization theory, the empirically documented pattern in which a location experiencing a crime, particularly burglary, faces substantially heightened risk of experiencing another crime within a short subsequent period and within close geographic proximity to the original incident, a pattern the Brantinghams’ crime pattern theory helps explain through offenders’ tendency to return to locations where they have already demonstrated successful criminal opportunity (Brantingham & Brantingham, 1993). This near-repeat pattern supplies predictive algorithms with a considerably shorter-term, more operationally actionable forecasting signal than the broader multi-year concentration stability Weisburd’s research documents, since near-repeat risk heightening typically decays within days or weeks following an initiating incident rather than persisting across the multi-year timeframes crime concentration research examines. Shane Johnson and Kate Bowers’s foundational empirical documentation of near-repeat burglary victimization patterns across multiple British cities established much of the original empirical basis for this near-repeat theoretical mechanism, finding consistent evidence that burglarized households and their immediate neighbors faced measurably heightened risk for a period of several weeks following an initial burglary, a finding subsequently replicated across American cities and incorporated directly into the commercial predictive policing algorithms this article examines (Johnson & Bowers, 2004).

This near-repeat theoretical foundation has proven particularly influential in shaping the specific algorithmic architecture predictive policing vendors have developed, since near-repeat models generate rapidly updating, short-term forecasts considerably more amenable to daily or weekly patrol deployment decisions than the more stable, longer-term hot spot designations traditional crime-mapping approaches, examined in the companion article on GIS and Crime Mapping elsewhere in this silo, typically generate. Rachel Boba Santos’s applied crime-analysis research further extended near-repeat theory’s practical application, developing specific tactical protocols for translating near-repeat risk forecasts into concrete patrol and notification procedures, including resident notification programs alerting neighbors of an initial burglary to their own temporarily heightened risk, a direct practical translation of near-repeat theory operating independently of the more algorithmically sophisticated commercial systems this article’s remaining sections examine (Santos, 2013).

Major Algorithmic Approaches

Place-Based Forecasting Systems

The most widely adopted category of predictive policing technology forecasts where crime is statistically likely to occur without attempting to predict who specifically will commit it, an approach that PredPol, among the most commercially prominent predictive policing vendors during the technology’s initial expansion period, popularized through an algorithm explicitly adapted from earthquake aftershock prediction models, reflecting the near-repeat victimization pattern’s structural similarity to the geographic and temporal clustering seismologists study in aftershock sequences (Perry, McInnis, Price, Smith, & Hollywood, 2013). Walter Perry and colleagues’ comprehensive RAND Corporation assessment of predictive policing technology found that these place-based forecasting systems generally built directly atop existing GIS crime-mapping infrastructure, adding a statistical forecasting layer to historical crime data already being routinely collected and mapped through the systems examined in the companion article on GIS and Crime Mapping elsewhere in this silo.

This place-based algorithmic approach has generally proven less controversial than person-based alternatives discussed below, since forecasting where crime is statistically likely to occur, without attempting to identify specific individuals as heightened risks, avoids some, though not all, of the due-process and civil-liberties concerns person-based predictive systems have generated. Andrew Ferguson’s broader legal analysis of predictive policing technology noted that even place-based systems raise meaningful Fourth Amendment and equal protection questions, since algorithmically designated hot spots can generate a form of individualized suspicion by association for anyone present within a predicted high-risk area, a legal concern distinct from, though related to, the more direct due-process implications person-based systems raise (Ferguson, 2017).

Person-Based Risk Assessment Systems

A second, considerably more controversial category of predictive policing technology attempts to identify specific individuals statistically likely to become involved in future violence, either as offenders or as victims, an approach the Chicago Police Department’s Strategic Subject List implementation made particularly prominent during the mid-2010s, generating a risk score for individuals based on factors including prior arrest history, gang affiliation, and social network proximity to previous shooting victims (Saunders, Hunt, & Hollywood, 2016). Jessica Saunders, Priscillia Hunt, and John Hollywood’s rigorous evaluation of the Chicago Strategic Subject List found that the algorithm’s predictive accuracy proved considerably weaker than departmental claims had suggested, and that the program’s actual operational implementation, involving police contact with identified high-risk individuals, showed no measurable effect on reducing those individuals’ involvement in future shootings either as victims or as perpetrators.

This person-based approach’s comparatively weak empirical performance, combined with its more direct civil-liberties implications given its focus on specific named individuals rather than geographic locations alone, has contributed to person-based predictive policing’s considerably more limited continued adoption relative to place-based systems, with several prominent early implementations, including Chicago’s Strategic Subject List, subsequently discontinued following critical evaluation and sustained public criticism. Sarah Brayne’s ethnographic research on predictive policing implementation within the Los Angeles Police Department documented how person-based risk lists, even where nominally intended to identify individuals warranting supportive social service intervention, functioned in practice primarily as tools intensifying routine police surveillance and contact with listed individuals, a gap between stated programmatic intent and actual operational effect that has fueled much of the sustained civil-liberties criticism person-based predictive systems have faced (Brayne, 2017).

Academic Research, Commercial Vendors, and International Comparison

Tensions Between Academic and Commercial Development

Predictive policing’s development has involved a distinctive and sometimes fraught relationship between academic criminologists, who developed much of the underlying near-repeat and crime concentration theory this article examines, and commercial technology vendors, who translated that academic research into proprietary, revenue-generating software products, a relationship that has generated persistent tension regarding intellectual credit, methodological transparency, and the appropriate boundary between rigorous academic research and commercial product marketing (Perry, McInnis, Price, Smith, & Hollywood, 2013). George Mohler and colleagues’ original academic development of the earthquake-aftershock-inspired predictive algorithm that PredPol subsequently commercialized illustrates this tension directly, since the underlying mathematical model originated in peer-reviewed academic research before being licensed and marketed as proprietary commercial technology whose specific operational parameters the commercial vendor subsequently declined to disclose in full (Mohler, Short, Brantingham, Schoenberg, & Tita, 2011).

This academic-commercial tension has shaped subsequent predictive policing research and policy debate considerably, since independent academic evaluation of commercial predictive policing products has often proven difficult given vendors’ proprietary restrictions on algorithm disclosure, complicating the kind of rigorous, independently replicable evaluation research that has characterized the hot spots policing evidence base examined in the companion article on Hot Spots Policing elsewhere in this silo.

International Adoption and Comparative Regulatory Approaches

Predictive policing technology, developed and initially commercialized primarily within the American policing context this article examines, has subsequently been adopted across a range of international policing systems, including departments in the United Kingdom, Netherlands, and Germany, generating a comparative regulatory landscape in which several European jurisdictions have implemented considerably more stringent algorithmic transparency and data-protection requirements than typically apply within the American regulatory context (Lum & Isaac, 2016). This comparative regulatory divergence has generated growing academic interest in whether European-style algorithmic accountability frameworks, including mandatory bias auditing and public disclosure requirements, might supply a productive model for addressing the transparency and feedback-loop concerns examined above within the American predictive policing context specifically.

This international comparison illustrates that the equity and transparency controversies examined throughout this article, while particularly prominent within American predictive policing debates given the technology’s early and widespread American commercial adoption, reflect a broader international challenge regarding algorithmic accountability in criminal justice contexts rather than a concern unique to the specific American departments and vendors this article has primarily examined.

Empirical Evidence on Forecasting Accuracy and Effectiveness

Testing Predictive Accuracy

Empirical research testing predictive policing algorithms’ actual forecasting accuracy has generally found modest rather than dramatic improvement over considerably simpler baseline approaches, including straightforward extrapolation from recent historical hot spot patterns, with several independent evaluations finding that sophisticated machine-learning algorithms offered only marginal accuracy gains over these simpler baseline methods despite requiring substantially greater technical infrastructure and cost to implement (Perry, McInnis, Price, Smith, & Hollywood, 2013). This modest accuracy advantage has generated genuine debate within the crime-analysis profession regarding whether predictive policing’s technical sophistication justifies its considerable implementation cost relative to simpler, more transparent, and more easily audited alternative approaches to place-based resource allocation.

Jerry Ratcliffe’s methodological research on hot spot identification offers a relevant comparative benchmark for interpreting this modest accuracy advantage, since his broader finding that different, comparatively simple statistical techniques for identifying spatial crime concentration can themselves generate meaningfully different practical conclusions suggests that predictive policing’s incremental accuracy improvements over simpler baseline methods may fall within the range of variation attributable to underlying methodological choice alone rather than representing a clearly decisive technological advance (Ratcliffe, 2004). Priscillia Hunt, Jessica Saunders, and John Hollywood’s broader RAND Corporation cost-effectiveness assessment of predictive policing technology extended this accuracy-skepticism into economic terms, finding that the considerable licensing and implementation costs many departments incurred adopting proprietary predictive policing software often exceeded what the technology’s modest documented incremental accuracy advantage over simpler alternatives would economically justify, a cost-effectiveness finding that has directly informed several of the program cancellation decisions examined later in this article (Hunt, Saunders, & Hollywood, 2014).

Crime Reduction Outcomes

Beyond pure forecasting accuracy, evaluations measuring predictive policing’s actual downstream effect on crime rates, rather than merely its forecasting accuracy in isolation, have produced decidedly mixed results, with some randomized field trials finding modest crime reductions in areas receiving predictive-informed patrol deployment while other evaluations found no statistically significant difference relative to conventional hot spots policing deployment based on simpler historical crime mapping alone (Perry, McInnis, Price, Smith, & Hollywood, 2013). This crime-reduction evidence gap, between predictive policing’s marketed technological sophistication and its actually documented incremental operational value beyond conventional hot spots approaches, has become a central point of contention in ongoing debates regarding whether continued predictive policing investment represents an efficient use of scarce departmental resources.

Anthony Braga’s broader meta-analytic research on hot spots policing effectiveness, examined in the companion article on Hot Spots Policing elsewhere in this silo, found consistently strong evidence for the underlying place-based resource-concentration strategy predictive policing extends, even as the specific incremental value predictive algorithms add beyond simpler historical hot spot identification remains considerably less firmly established (Braga, Papachristos, & Hureau, 2014). David Weisburd’s broader assessment of the crime concentration research underlying predictive policing’s theoretical claims noted that this modest incremental value finding should not be entirely surprising given the underlying statistical logic, since if crime concentration already displays the substantial year-over-year stability his own research documents, simple extrapolation from recent historical patterns should already capture most of the predictable signal any more sophisticated algorithm could hope to extract (Weisburd, 2015).

Equity and Civil Liberties Controversies

The Feedback Loop and Historical Bias Concern

Predictive policing’s most persistent and theoretically serious critique concerns the risk that algorithms trained on historical crime data will reproduce and potentially amplify whatever reporting and enforcement biases already exist within that underlying data, since neighborhoods experiencing historically heavier police presence generate correspondingly more recorded crime data through greater detection opportunity alone, creating a feedback loop in which historical enforcement patterns generate predictions that justify continued or intensified enforcement in the same areas independent of any genuine underlying difference in criminal activity (Lum & Isaac, 2016). Kristian Lum and William Isaac’s influential technical analysis of this feedback-loop concern, applying predictive policing algorithms to synthetic data with known ground-truth crime patterns, found that algorithmic predictions systematically diverged from actual underlying crime distribution in ways consistent with the historical-bias amplification concern, providing empirical rather than purely theoretical support for this critique. Solon Barocas and Andrew Selbst’s broader legal and technical analysis of algorithmic disparate impact supplied additional theoretical grounding for the feedback-loop concern Lum and Isaac’s empirical work documented, arguing that machine-learning systems trained on historically biased data will, absent deliberate corrective intervention, systematically reproduce that historical bias regardless of the algorithm designers’ intentions, a general principle with direct application to the predictive policing feedback-loop mechanism this section examines (Barocas & Selbst, 2016).

This feedback-loop concern connects directly to the data-quality issues examined in the companion article on GIS and Crime Mapping elsewhere in this silo, since predictive policing represents merely the most operationally consequential application of a data-quality concern that applies, in principle, to any crime-mapping and analysis system relying on officially recorded rather than independently and objectively measured crime data.

Transparency and Due Process Concerns

Beyond the feedback-loop critique, predictive policing has generated sustained concern regarding algorithmic transparency, since many commercial predictive policing systems operate as proprietary technology whose specific underlying algorithms and weighting factors remain undisclosed even to the police departments purchasing and deploying them, complicating independent audit and raising due-process concerns for any individual whose treatment by police may be influenced by an algorithmic prediction they cannot meaningfully examine or contest (Saunders, Hunt, & Hollywood, 2016). This transparency concern has motivated growing legislative and policy attention to algorithmic accountability requirements specifically targeting predictive policing and related criminal justice risk-assessment technology, with several state and municipal jurisdictions adopting disclosure and audit requirements as a condition of continued predictive policing use. Danielle Citron and Frank Pasquale’s broader legal scholarship on algorithmic accountability in scoring systems supplied an influential doctrinal framework for these emerging disclosure requirements, arguing that any consequential algorithmic scoring system affecting individual liberty or opportunity, predictive policing risk scores among them, should be subject to meaningful procedural due process protections including the right to know the specific factors contributing to one’s own score and the opportunity to contest inaccurate underlying data (Citron & Pasquale, 2014).

Cynthia Lum and Christopher Koper’s broader evidence-based policing assessment situated these transparency concerns within a wider evaluation of predictive policing’s overall evidentiary and practical standing, concluding that the technology’s combination of modest demonstrated incremental effectiveness and substantial equity and transparency concerns positioned it considerably less favorably within the evidence-based policing framework than the more established hot spots policing strategies examined in the companion article on Hot Spots Policing elsewhere in this silo (Lum & Koper, 2017).

Program Cancellations and the Field’s Contemporary Trajectory

High-Profile Discontinuations

The mid-to-late 2010s and early 2020s saw a wave of prominent predictive policing program cancellations across major American departments, including the Los Angeles Police Department’s discontinuation of its PredPol implementation and the Chicago Police Department’s discontinuation of its Strategic Subject List, decisions substantially influenced by the combination of modest documented effectiveness, sustained civil-liberties advocacy, and, in several cases, independent audits finding limited evidence of the crime-reduction benefits departments had initially anticipated (Saunders, Hunt, & Hollywood, 2016). The Los Angeles Police Department’s own inspector general audit, commissioned following sustained community advocacy, found that the department’s predictive policing implementation had been marked by inconsistent operational adherence to the underlying algorithmic recommendations and inadequate ongoing evaluation of the program’s actual crime-reduction effectiveness, procedural findings that reinforced the broader evaluation research examined earlier in this article and directly informed the department’s subsequent discontinuation decision (Los Angeles Police Department Office of the Inspector General, 2019). These high-profile cancellations have generated broader reconsideration within the American policing profession regarding predictive policing’s appropriate role, with many departments that previously expressed enthusiasm for the technology adopting a more cautious, evidence-demanding posture toward continued or renewed predictive policing investment.

This cancellation wave does not represent complete abandonment of algorithmic and statistical forecasting within American policing, since many departments have continued to employ simpler, more transparent statistical forecasting approaches, including straightforward historical hot spot extrapolation, while specifically discontinuing the more sophisticated, proprietary, and less transparent machine-learning systems that generated the most sustained civil-liberties controversy.

Emerging Alternative Approaches

Contemporary departments reconsidering predictive policing have increasingly gravitated toward more transparent, publicly auditable forecasting approaches, including open-source statistical models whose underlying methodology remains available for independent academic and civil-society review, a transparency-oriented shift that directly addresses the algorithmic accountability concerns examined above while preserving some of the forecasting value the broader predictive policing concept originally promised (Lum & Isaac, 2016). This emerging transparency-oriented approach has also incorporated more explicit attention to potential historical bias within training data, with some contemporary systems specifically incorporating bias-auditing and correction procedures designed to address the feedback-loop concern Lum and Isaac’s research identified. Rashida Richardson, Jason Schultz, and Kate Crawford’s research on predictive policing data integrity extended this bias-auditing concern by documenting several specific cases in which departments continued relying on predictive policing systems trained partly on data generated during periods of documented unlawful or unconstitutional policing practice, a data-provenance concern distinct from but related to the more general historical-bias mechanism Lum and Isaac examined, and one that has motivated some jurisdictions to require formal data-integrity certification before permitting continued predictive policing use (Richardson, Schultz, & Crawford, 2019).

This ongoing methodological evolution illustrates that predictive policing’s underlying technical and theoretical ambition, forecasting crime concentration with greater precision than simpler historical extrapolation alone provides, remains an active area of continued development even as the specific proprietary, opaque implementations that generated the most severe civil-liberties controversy during the technology’s initial 2010s expansion have been substantially curtailed or discontinued across the American departments that had most enthusiastically adopted them.

Conclusion

Predictive policing in America represents environmental criminology’s most technologically ambitious and most controversial extension of the place-based crime concentration research examined throughout this silo, translating Weisburd’s documented crime concentration stability and the Brantinghams’ near-repeat victimization theory into operational forecasting tools deployed across numerous American police departments beginning around 2010. This technological ambition has generated only modest empirically documented improvement in forecasting accuracy and crime-reduction effectiveness relative to considerably simpler historical hot spot extrapolation methods, a modest evidentiary return that, combined with serious feedback-loop and transparency concerns Lum and Isaac’s research and sustained civil-liberties advocacy have documented, has motivated a substantial wave of program discontinuation across major American departments during the mid-to-late 2010s and beyond.

This trajectory, from initial enthusiastic adoption through critical evaluation to substantial program cancellation and cautious methodological reconsideration, illustrates environmental criminology’s continued capacity for genuine empirical self-correction even regarding its own most technologically sophisticated applications, confirming that the field’s commitment to rigorous evaluation, documented throughout this silo’s treatment of hot spots policing and displacement research, extends fully to predictive policing’s more recent and more contested technological innovations. Francis Cullen’s broader assessment of evidence-based criminal justice reform noted that this predictive policing trajectory, unlike many previous criminal justice technology adoptions that persisted for decades despite weak evidentiary support, demonstrates an unusually rapid empirical accountability cycle, with independent academic and civil-society evaluation exerting measurable influence on departmental technology decisions within a relatively compressed timeframe of roughly one decade from initial widespread adoption to substantial reconsideration (Cullen, 2011).

References

Braga, A. A., Papachristos, A. V., & Hureau, D. M. (2014). The effects of hot spots policing on crime: An updated systematic review and meta-analysis. Justice Quarterly, 31(4), 633–663.
Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104(3), 671–732.
Brantingham, P. J., & Brantingham, P. L. (1993). Nodes, paths and edges: Considerations on the complexity of crime and the physical environment. Journal of Environmental Psychology, 13(1), 3–28.
Brayne, S. (2017). Big data surveillance: The case of policing. American Sociological Review, 82(5), 977–1008.
Citron, D. K., & Pasquale, F. (2014). The scored society: Due process for automated predictions. Washington Law Review, 89(1), 1–33.
Cohen, L. E., & Felson, M. (1979). Social change and crime rate trends: A routine activity approach. American Sociological Review, 44(4), 588–608.
Cullen, F. T. (2011). Beyond adolescence-limited criminology: Choosing our future. Criminology, 49(2), 287–330.
Ferguson, A. G. (2017). The rise of big data policing: Surveillance, race, and the future of law enforcement. New York University Press.
Groff, E. R., Weisburd, D., & Yang, S. M. (2010). Is it important to examine crime trends at a local “micro” level? A longitudinal analysis of street to street variability in crime trajectories. Journal of Quantitative Criminology, 26(1), 7–32.
Hunt, P., Saunders, J., & Hollywood, J. S. (2014). Evaluation of the Shreveport predictive policing experiment. RAND Corporation.
Johnson, S. D., & Bowers, K. J. (2004). The burglary as clue to the future: The beginnings of prospective hot-spotting. European Journal of Criminology, 1(2), 237–255.
Lum, C., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.
Los Angeles Police Department Office of the Inspector General. (2019). Review of selected Los Angeles Police Department data-driven policing strategies. Los Angeles Police Commission.
Lum, C., & Koper, C. S. (2017). Evidence-based policing: Translating research into practice. Oxford University Press.
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.
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.
Ratcliffe, J. H. (2004). The hotspot matrix: A framework for the spatio-temporal targeting of crime reduction. Police Practice and Research, 5(1), 5–23.
Richardson, R., Schultz, J. M., & Crawford, K. (2019). Dirty data, bad predictions: How civil rights violations impact police data, predictive policing systems, and justice. New York University Law Review Online, 94, 15–55.
Santos, R. B. (2013). Crime analysis with crime mapping (3rd ed.). Sage Publications.
Saunders, J., Hunt, P., & Hollywood, J. S. (2016). Predictions put into practice: A quasi-experimental evaluation of Chicago’s predictive policing pilot. Journal of Experimental Criminology, 12(3), 347–371.
Weisburd, D. (2015). The law of crime concentration and the criminology of place. Criminology, 53(2), 133–157.