Hot Spots, Feedback Loops, and Vendor Claims: A Technical Audit of Place-Based Predictive Policing

Hot Spots, Feedback Loops, and Vendor Claims: A Technical Audit of Place-Based Predictive Policing

Executive findings and scope

This report audits place-based predictive policing as a sociotechnical intervention, not merely as a forecasting algorithm. The relevant causal chain is:

\[ \text{historical data} \rightarrow \text{risk estimate} \rightarrow \text{patrol assignment} \rightarrow \text{officer behavior} \rightarrow \text{crime, reporting, and enforcement outcomes} \rightarrow \text{future data}. \]

A technically accurate forecast can fail operationally because officers do not visit the forecast area, because the recommended intervention is ineffective, or because the outcome measure does not distinguish deterrence from reporting and recording changes. Conversely, conventional hot-spot patrol can reduce crime even when an algorithm contributes little or nothing beyond recent-event counts, kernel-density maps, or an experienced analyst. The National Institute of Justice’s Shreveport evaluation explicitly concluded that it remained unclear whether predictive maps were more useful than traditional maps of prior crime locations; the study found no statistically significant incremental reduction in property crime. citeturn21view0turn21view1

The necessary technical distinctions

The following categories should not be collapsed into one label merely because they use maps, databases, artificial intelligence, or centralized police technology.

A map can therefore be nonpredictive, and a product marketed as “resource management,” “precision policing,” “mission planning,” or “risk reduction” can remain predictive if it ranks future place-time risk and directs patrol accordingly. Functional behavior, not branding, is the controlling criterion.

Principal conclusions

The strongest evidence supports focused policing at genuinely high-crime microplaces, not the proposition that proprietary predictive software substantially improves on transparent hot-spot methods. A Campbell systematic review covering 65 studies and 78 tests found a small but statistically significant crime-control effect from hot-spots policing; displacement was not inevitable, and nearby areas sometimes experienced diffusion of benefits. Those results establish the value of concentration and focused intervention, but they do not establish the incremental value of PredPol, HunchLab, RTM software, or machine learning. citeturn3search2turn3search6turn22search11

The best-known favorable PredPol field experiment found that an epidemic-type aftershock sequence model forecast between 1.4 and 2.2 times as much crime as analyst-produced maps and estimated a 7.4 percent reduction in crime as a function of patrol time in Los Angeles treatment conditions. The experiment used day-level random assignment and visually identical treatment and control maps. It is an important result, but several authors were PredPol co-founders or shareholders, the analysis intertwined model selection and patrol delivery, and the article itself called for better GPS measurement and research on the feedback between patrol, crime, and algorithms. citeturn18view2turn19view0turn19view1turn27view0

An independently funded Philadelphia randomized trial of HunchLab found a 31 percent reduction in expected property-crime counts for a dedicated marked-car intervention, with an estimated 40 percent reduction during the following eight-hour period. Mere awareness of forecast areas and a dedicated unmarked-car strategy did not show comparable benefits; no intervention produced a persuasive violent-crime reduction. The result is better interpreted as evidence about visible patrol dosage in selected microplaces than as proof that HunchLab’s proprietary forecasting added value over a simple baseline, because the experiment did not randomize HunchLab forecasts against recent-event or conventional analyst hot spots. citeturn18view3turn19view2turn27view2turn27view3

Independent evidence from Plainfield, New Jersey, is far less favorable to PredPol/Geolitica. Of 5,722 robbery or aggravated-assault forecasts without recorded patrol dosage, 32 forecast records corresponded to 20 of 139 reported incidents. That implies alert-level precision of 0.56 percent and incident-level recall of 14.4 percent. Only 8 of 116 burglaries occurred in the predicted place-time windows, for 6.9 percent recall; the published alert success rate was approximately 0.1 percent. Officers entered forecast boxes during only 129 of 23,760 forecast instances, and department officials said the system was rarely, if ever, used to direct patrol. Thus the case is simultaneously an unfavorable forecast evaluation and an implementation failure—not a valid test of whether faithful PredPol patrol reduces crime. citeturn26view0turn26view1turn27view5

The Shreveport Predictive Intelligence-Led Operational Targeting experiment likewise found no statistically significant incremental crime reduction. Treatment districts reportedly spent 6–10 percent less than controls, largely because of lower overtime, but heterogeneous implementation, few districts, low event counts, and limited statistical power prevented a strong conclusion. RAND recommended a direct comparison in which treatment and control receive equivalent interventions and differ only in whether locations come from an algorithm or traditional crime mapping. citeturn21view0turn21view1turn21view2

Risk-terrain modeling has credible evidence that environmental features can identify places with elevated future risk. A systematic review found 25 studies and reported that roughly half of future cases were captured in the highest-risk 10 percent of cells across the pooled literature. That is evidence of spatial discrimination, not automatically of calibration or causal crime reduction. Promotional claims associated with Atlantic City and Rutgers frequently cite reductions around 20–30 percent, but the prominent Atlantic City comparisons were before-and-after figures accompanied by business checks, lighting changes, vacant-building interventions, and other measures. Without a randomized or credible matched counterfactual, those changes cannot be attributed uniquely to RTM. citeturn22search1turn22search4turn22search20turn22search28

Vendor consolidation has not eliminated the underlying functions. Azavea sold HunchLab to ShotSpotter in 2018; it became ShotSpotter Missions and later ResourceRouter. PredPol renamed itself Geolitica in 2021 and ceased operating independently at the end of 2023, after SoundThinking hired its engineering team and acquired technology, intellectual property, and customer-related assets. SoundThinking’s 2025 annual filing described HunchLab as renamed ResourceRouter and said the product supports strategically planned patrols. Geolitica’s corporate exit therefore did not mark the disappearance of predictive patrol allocation; it marked an acquisition and migration into a broader policing platform. citeturn17search1turn17search3turn17search6turn17search20turn18view4turn26view0

The central accountability problem is not only algorithmic opacity. It is the frequent absence of preserved forecast outputs, patrol-dosage records, denominator definitions, baseline models, preregistered evaluation plans, contract deliverables, and auditable links between forecast, dispatch, officer presence, enforcement activity, and subsequent outcomes. NYPD records, for example, indicated that its in-house predictive tool did not retain or recreate historical outputs, making retrospective validation substantially more difficult. citeturn24view1

Technical taxonomy and vendor genealogy

Traditional hot spots, near repeats, and self-excitation

Crime is spatially concentrated, but concentration has at least three analytically different components.

The first is persistent heterogeneity: some street segments, facilities, and small areas have chronically higher recorded crime than others because of land use, opportunity structures, traffic, guardianship, population exposure, and enduring social conditions. A historical count, moving average, or kernel-density estimate can exploit this persistence without making assumptions about contagion.

The second is repeat victimization: a previously victimized address or target experiences temporarily increased risk. The third is near-repeat concentration: after an event, nearby places experience elevated risk for a limited period. Prospective hot-spotting operationalizes these patterns by creating temporary risk buffers or surfaces around recent incidents. Research supports near-repeat structure for some property offenses, particularly burglary and vehicle crime, but predictive usefulness varies by offense, geography, event volume, spatial threshold, and forecast horizon. Short-lived near-repeat signals can also be operationally difficult because a forecast may expire before an agency can deliver a meaningful preventive response. citeturn22search2turn22search22turn22search26turn22search29

A self-exciting point process formalizes this idea. In simplified form, the conditional event intensity at location \(s\) and time \(t\) is:

\[ \lambda(s,t) = \mu(s) + \sum_{i:t_i<t} g(s-s_i,t-t_i), \]

where \(\mu(s)\) represents persistent background risk and \(g(\cdot)\) represents the temporary increase associated with prior events. PredPol’s intellectual ancestry lies in earthquake aftershock models: an incident is treated analogously to a main shock that may be followed by nearby events, although the criminological mechanisms can include repeat targeting, offender foraging, shared opportunity, reporting artifacts, and police detection rather than literal contagion. The foundational crime-modeling article was published by Mohler and colleagues in 2011. citeturn0search14turn18view2

The distinction between background and triggering components is useful, but it does not solve the label problem. The process models recorded events, not latent criminal behavior directly. If patrol increases drug arrests in a forecast area, the resulting event pattern may be self-exciting in the database even when underlying drug use is geographically similar elsewhere. If residents in one area report burglary more often, the fitted background rate can reflect reporting access and insurance requirements as well as victimization.

PredPol and Geolitica

PredPol emerged from collaboration among researchers associated with UCLA and police agencies in Los Angeles and Santa Cruz. Its classic implementation used a deliberately narrow input schema—incident type, location, and time, plus identifiers and modification timestamps for data management—and produced ranked small boxes indicating elevated near-term risk. Records disclosed during NYPD’s vendor evaluation described crime type, geographic coordinates, occurrence time, incident identifier, and optional record-update time as the core fields. citeturn24view2

The standard public description emphasized roughly 500-by-500-foot prediction boxes, refreshed for daily or shift-level patrol. The product’s appeal was operational simplicity: officers received a small number of red boxes and were encouraged to enter them when not handling calls. Some implementations recorded “dosage,” meaning time spent inside a forecast box, through vehicle location, application interaction, or dispatcher records. That dosage was central to the vendor’s causal story: a forecasted event that did not occur after patrol could be counted as possible deterrence rather than a false positive.

That logic makes dosage indispensable to evaluation. It is impossible to interpret the absence of a subsequent incident without knowing whether treatment was delivered. Yet The Markup identified prediction data for 38 jurisdictions and obtained meaningful long-duration dosage records from only Plainfield; most agencies said the records were unavailable or nonpublic. citeturn27view5

PredPol renamed itself Geolitica in 2021. Geolitica’s independent operations ended in 2023, and customers were offered migration to SoundThinking’s platform. Contemporary accounts distinguished among the engineering employees, patents or other intellectual property, customer relationships, and the original source code; SoundThinking stated that it had not acquired the old crime-prediction source code, while its securities filings described acquisition of Geolitica’s primary technology and intellectual property. The most defensible conclusion is that the corporate entity and original product ended, while selected personnel, intellectual property, and customer functions were absorbed into a successor patrol-allocation ecosystem. citeturn17search3turn17search6turn17news23turn26view0

HunchLab, Missions, and ResourceRouter

HunchLab began with a Philadelphia “Crime Spike Detector,” developed with the police department and the U.S. Attorney’s Office. Azavea received National Science Foundation small-business funding and expanded the prototype between approximately 2008 and 2011. Early functions included localized spike detection, seasonal crime-load forecasting, and near-repeat methods. HunchLab 2.0 later combined multiple feature families and a machine-learning forecasting model. citeturn18view5turn17search20

Its documented inputs were substantially broader than PredPol’s. They could include historical crime at several look-back windows; recent three-, seven-, and fourteen-day patterns; zoning and land-use variables; schools, hospitals, transit, roads, water, and public-safety facilities; income, rent, housing, vacancy, population, and vehicle measures; temporal cycles; events; and weather. In the Philadelphia experiment, HunchLab used gradient-boosted decision trees, cross-validation, and smoothing to convert model scores into expected counts for 500-by-500-foot cells. The highest-scoring cells became “mission grids.” The specific variable importance logs were no longer available to the evaluators after the ownership transfer, illustrating how acquisition can sever the evidentiary chain needed for later audits. citeturn27view2

HunchLab also went beyond forecasting into mission generation. It ranked expected crime per unit of patrol effort, considered available officers and vehicles, and distributed printable or mobile missions. Its promotional interface contemplated tactics such as neighborhood patrol, vehicle stops, reviewing known offenders, or interviewing residents. The inclusion of tactic recommendations is important: two agencies could use identical forecast scores but generate very different community impacts depending on whether officers provide visible presence, engage businesses, conduct stops, seek arrests, or coordinate environmental remediation. citeturn18view5turn19view4

ShotSpotter acquired HunchLab in 2018 and marketed it as ShotSpotter Missions. SoundThinking’s later filings identify ResourceRouter as the renamed descendant. The genealogy is therefore:

\[ \text{Crime Spike Detector} \rightarrow \text{HunchLab} \rightarrow \text{ShotSpotter Missions} \rightarrow \text{ResourceRouter}. \]

The names changed, and product components evolved, but the continuing function—estimating risk and recommending patrol missions—places the forecasting and allocation modules within the predictive-policing audit boundary. citeturn17search1turn17search10turn18view4

Risk-terrain modeling

Risk-terrain modeling begins from the environmental-criminology proposition that crime opportunities are produced partly by configurations of places: bars, transit nodes, vacant properties, street networks, commercial facilities, poorly managed parcels, and other features may attract targets, facilitate offending, or weaken guardianship. RTM creates standardized spatial layers representing proximity, density, or exposure and combines selected factors into a risk surface. Its original shooting applications were explicitly presented as forecasts of where future shootings would be distributed. citeturn22search12turn22search16turn22search31

RTM differs from event-driven systems in three respects. First, its features are often more stable than yesterday’s incident pattern. Second, its intended intervention can be environmental—lighting, code enforcement, property management, business engagement, or removal of an opportunity structure—rather than repeated patrol. Third, the model can have an explanatory or diagnostic purpose even when it does not generate short-term probability forecasts.

RTM’s strengths are transparency of spatial factors and potential linkage to nonenforcement remedies. Its weaknesses include researcher degrees of freedom in selecting features, distances, operationalizations, and model-selection criteria; risk of encoding segregated land-use patterns; and frequent reliance on observational validation. Capturing many future incidents in the top risk decile establishes ranking ability, but not necessarily well-calibrated probabilities or a causal effect from acting on the map. citeturn22search4turn22search28

Comparable police-built systems

Police-built systems deserve the same functional scrutiny as commercial products. NYPD integrated predictive algorithms into its Domain Awareness System after earlier hand-built and computer-generated hot-spot mapping. From 2015 to 2016, it evaluated PredPol, HunchLab, and Keystats in a no-cost trial requiring five small forecast boxes by precinct, platoon, and crime category. PredPol withdrew, and NYPD ultimately used an internally developed algorithm. The department’s refusal and delayed production of records—and the reported inability to recreate old outputs—made independent evaluation difficult. citeturn23search2turn24view1turn24view2

NYPD’s Patternizr, by contrast, should not be treated as a place forecast. It is an investigative pattern-matching tool for connecting robbery, burglary, and larceny reports. Its use of machine learning and maps does not convert it into a system predicting future place-time risk.

LAPD’s Operation LASER also illustrates category overlap. LASER combined chronic-location analysis with person-focused “chronic offender” mechanisms. Its place component belongs in a place-based audit, but its individual scoring must be separately assessed under person-based due-process and bias standards. LAPD’s inspector general reviewed LASER, PredPol, and related data-driven practices and found inadequate documentation and insufficient evidence to isolate their crime effects. LAPD ended LASER in 2019. citeturn1search13turn1search17turn1search25turn23search13

Shreveport’s PILOT program is another useful comparator because it was built around agency analytics rather than a major commercial license. It combined historical crime locations, calls, field interviews, and prior hot spots to identify small property-crime risk areas, then directed locally selected prevention activities. Its evaluation demonstrates that in-house status does not ensure effectiveness, and proprietary status is not the defining issue. citeturn21view1turn21view2

Deployment audit and evidence matrix

The tables below distinguish documented facts from missing information. “Not publicly located” is not equivalent to zero cost, no training window, or no validation; it means the available record did not support a reliable entry. Operational statuses are stated as of August 2, 2026, where records permit.

City-by-city technical and operational table

Procurement, validation, code, and status table

Independent and vendor-connected evidence matrix

Methodological audit and public-data reproduction

Forecast metrics and their denominators

Predictive-policing claims frequently use “accuracy” without identifying the unit of analysis. At least five quantities can be called a success rate:

\[ \text{Incident hit rate} = \frac{\text{unique future incidents inside forecast areas}} {\text{all future incidents}} \]

\[ \text{Alert precision} = \frac{\text{forecast records matched to an incident}} {\text{all forecast records}} \]

\[ \text{Cell precision} = \frac{\text{forecast cells containing at least one incident}} {\text{forecast cells issued}} \]

\[ \text{Coverage} = \frac{\text{forecasted geographic area}} {\text{eligible study area}} \]

\[ \text{Prediction Accuracy Index} = \frac{\text{incident hit rate}}{\text{coverage}}. \]

The National Institute of Justice used PAI in its Real-Time Crime Forecasting Challenge. Entries forecast between 0.25 and 0.75 square miles within a 147.71-square-mile Portland study area, and PAI rewarded the proportion of events captured relative to the proportion of area designated. citeturn26view3turn8view0

PAI is useful for comparing forecasts at similar scales, but it is not a probability-accuracy measure. A tiny forecast area can obtain a large PAI from very few events. PAI also changes if the denominator includes parks, industrial land, water, highways, or areas with no meaningful population exposure. Street length, address count, businesses, households, resident population, ambient population, or patrol-accessible area may be more appropriate denominators for particular questions. Published methodological critiques show that PAI rankings can depend materially on these choices. citeturn3search8turn3search19

Precision and recall answer different operational questions. Precision measures forecast burden: how many issued alerts correspond to the target outcome? Recall measures coverage of harm: how many target incidents were captured? A system can achieve high recall by designating much of a city, but that produces low precision and little prioritization. A system can achieve high PAI by designating a tiny area around one persistent location while missing most citywide incidents.

Calibration asks whether predicted probabilities correspond to observed frequencies. For cells assigned risk \(p\), approximately \(p\) of equivalent cell-periods should contain an event. Calibration can be evaluated with reliability plots, calibration intercept and slope, Brier score, and log loss. A product that releases only ranked red boxes without probabilities cannot be fully calibrated by outsiders. “Twice average risk” is not interchangeable with “a 20 percent probability.”

Forecast lead time is the gap between model issuance and the target period. A forecast made after a late report may inadvertently include information unavailable at the intended decision time. Data pipelines must use immutable “known as of” timestamps, not merely occurrence dates, or they risk leakage from revised reports, delayed entry, or retrospective geocoding.

Baseline comparison must include more than random allocation. At minimum, a proposed system should be compared with:

\[ \begin{aligned} B_1 &: \text{recent event count by cell},\\ B_2 &: \text{historical average by cell and season},\\ B_3 &: \text{kernel-density estimate},\\ B_4 &: \text{last-period persistence},\\ B_5 &: \text{analyst-selected hot spots},\\ B_6 &: \text{same patrol budget allocated to conventional hot spots}. \end{aligned} \]

The Los Angeles/Kent publication did include analysts and three- and seven-day count maps. Analysts were statistically indistinguishable from seven-day count maps in the reported Kent comparisons, while ETAS performed better. This is precisely the type of transparent benchmark needed, although it should be independently replicated on multiple cities and locked evaluation periods. citeturn18view2turn19view1

Reproduction of the Plainfield evaluation

The Markup’s public repository contains prediction data, dosage data, census-block-group demographics, preparation scripts, an analysis notebook, dependency files, and a make reproduce workflow. It joins Geolitica forecasts to Plainfield police reports and calculates success rates. This is unusually strong reproducibility practice for a police-technology evaluation. citeturn26view2

The arithmetic below reproduces core published metrics from the released denominators. It is a denominator-level reproduction rather than a claim to have re-geocoded every address independently.

Robbery and aggravated assault

Published counts:

\[ N_{\text{alerts}} = 5{,}722 \]

\[ N_{\text{matched alert records}} = 32 \]

\[ N_{\text{actual incidents}} = 139 \]

\[ N_{\text{unique incidents captured}} = 20. \]

Therefore:

\[ \text{Alert precision} = \frac{32}{5{,}722} = 0.005592 = 0.559\%. \]

\[ \text{Incident recall} = \frac{20}{139} = 0.1439 = 14.4\%. \]

The alert workload was:

\[ \frac{5{,}722}{20} = 286.1 \]

forecast records per unique captured robbery or aggravated assault. Because overlapping predictions sometimes matched the same event, alert precision and incident recall cannot be inferred from one another. citeturn27view5

Burglary

Published counts:

\[ N_{\text{alerts}} = 10{,}141, \quad N_{\text{actual burglaries}} = 116, \quad N_{\text{captured burglaries}} = 8. \]

Thus:

\[ \text{Incident recall} = \frac{8}{116} = 6.90\%. \]

There were approximately:

\[ \frac{10{,}141}{8} = 1{,}267.6 \]

burglary forecast records for each unique burglary captured. The publication reported alert success of approximately 0.1 percent; the exact number of matched burglary forecast records, as distinct from unique incidents, is required for a more precise alert-level calculation. citeturn27view5

Overall target-incident coverage and patrol delivery

Across the forecast-eligible target categories, the public methodology reported 34 of 336 relevant incidents captured:

\[ \text{Overall incident recall} = \frac{34}{336} = 10.12\%. \]

After treated prediction records were excluded, 23,631 forecast records remained:

\[ \frac{23{,}631}{336} = 70.3 \]

untreated forecast records per relevant incident occurring anywhere in the city during the period.

Actual patrol occurred during 129 of 23,760 forecast instances:

\[ \text{Observed forecast-box dosage rate} = \frac{129}{23{,}760} = 0.543\%. \]

Dosage records averaged about 20 minutes, although the agency stated that those overlaps were coincidental rather than forecast-directed. citeturn26view1turn27view5

These calculations support three conclusions.

First, Plainfield’s forecast output had very low alert precision and low incident recall under The Markup’s matching rules. Second, the department delivered almost no forecast-linked treatment, so the procurement produced negligible operational exposure. Third, the case cannot resolve whether a properly implemented PredPol patrol strategy would reduce crime, because the treatment mechanism was absent.

A full independent replication should additionally recalculate results under alternative spatial buffers, exact versus interval-censored occurrence times, address-geocoding uncertainty, offense-classification rules, unique cell-period rather than record-level precision, and simple recent-count baselines. The original methodology already used a generous 600-by-600-foot matching box around nominal 500-foot forecasts and documented ambiguous addresses, report-time uncertainty, and offense hierarchy limitations. citeturn15view0turn26view1

Forecast accuracy is not crime reduction

A forecast evaluation observes whether future recorded events fall inside designated places. A crime-reduction evaluation asks whether acting on the forecast changes outcomes relative to a credible counterfactual.

Let \(F_j\) denote the forecast method, \(D_j\) patrol dosage, \(T_j\) the tactics delivered, and \(Y_j\) subsequent crime. A simple causal structure is:

\[ F_j \rightarrow D_j \rightarrow T_j \rightarrow Y_j. \]

But the same forecast can produce different \(D_j\) and \(T_j\), while call demand, staffing, weather, neighborhood conditions, and existing patrol priorities affect both dosage and outcomes. Comparing high-dosage boxes with low-dosage boxes is confounded because officers may spend more time where risk is visibly greater.

The cleanest factorial evaluation would randomize both forecast source and intervention:

The contrast between proprietary-model patrol and transparent-baseline patrol estimates algorithmic value only if patrol time, tactics, staffing, forecast-area size, offense definition, and temporal coverage are held constant. Shreveport failed to maintain fully comparable interventions across districts, which is why RAND could not cleanly isolate map quality. citeturn21view0

Philadelphia showed a related principle: awareness alone was insufficient, while a dedicated marked car produced a property-crime effect. The active ingredient appears to have included visible, sustained patrol rather than the simple existence of a risk score. citeturn18view3turn19view2

Displacement, diffusion, and contamination

A treatment-area decline is not enough. Offending may move to adjacent places or later times:

\[ \text{Net effect} = \Delta Y_{\text{treatment}} + \Delta Y_{\text{buffer}} + \Delta Y_{\text{rest of jurisdiction}}. \]

Spatial displacement occurs if crime shifts into nearby untreated areas. Temporal displacement occurs if it shifts beyond the treated shift. Target or offense displacement occurs if offenders switch victims or offense categories. Diffusion of benefits occurs if deterrence extends beyond the treated area or time.

The Philadelphia experiment reported a reduction during the subsequent eight-hour period after marked-car property-crime patrol, consistent with temporal diffusion. Its evaluators also expanded the treatment unit to adjacent cells because officers could not realistically remain inside one 500-foot square and had to travel along surrounding streets. That adjustment illustrates why nominal forecast geometry is not the same as the actual treatment footprint. citeturn18view3turn27view2

Treatment contamination arises when control officers enter treatment areas, commanders independently target the same places, citywide operations affect both groups, or forecast maps reveal information used outside assigned conditions. GPS or automated vehicle-location data should measure all officer presence, not only self-declared “missions.” The Los Angeles study acknowledged that call logs were less precise than in-car GPS for this purpose. citeturn27view0

Patrol dosage and dose response

Dosage should be measured as more than a binary visit:

\[ D_{it} = \sum_o \left( \text{officer-minutes}_{oit} \times \text{visibility}_{oit} \times \text{tactic weight}_{oit} \right). \]

Required components include officer-minutes, number of officers, marked versus unmarked presence, stationary versus moving patrol, pedestrian versus vehicle presence, stops, searches, business contacts, arrests, calls handled, and interruptions. “Entered box” is not equivalent to meaningful treatment.

Analysis should estimate a dose-response curve and test diminishing returns. Repeated fifteen-minute patrol episodes may differ from one uninterrupted hour. High dosage can deter visible street offenses while increasing recorded drug, weapons, traffic, and disorder enforcement. Outcomes must therefore distinguish victim-reported offenses from police-initiated detections.

Data generation, feedback loops, and racial geography

Recorded crime is a selected measurement process

Underlying victimization is not directly observed in police records. A simplified measurement model is:

\[ R_{ist} = U_{ist} \times P(\text{reported}\mid U,X_{ist}) \times P(\text{recorded}\mid \text{report},Z_{ist}), \]

where \(U\) is underlying victimization, \(X\) includes victim trust, offense severity, insurance requirements, language access, immigration concerns, and police accessibility, and \(Z\) includes agency recording rules, classification, workload, and discretion.

The National Crime Victimization Survey measures both reported and unreported nonfatal victimization, whereas police incident systems measure crimes known to law enforcement. In 2024, only about three in ten property victimizations were reported to police, including approximately 41 percent of burglaries. For 2020–2023, BJS estimated that about 38 percent of urban violent victimizations were reported, compared with 43 percent in suburban and 51 percent in rural areas; reporting also varied sharply by offense. citeturn25search1turn25search5turn25search16turn25search17turn25search20

Reported burglary is generally closer to a victim-driven measure than a narcotics arrest, but it remains selected. Residents may report because of insurance, property value, confidence that police will respond, surveillance-camera availability, or institutional requirements. Conversely, fear of retaliation, prior negative police contact, informal resolution, or belief that police cannot help can suppress reporting.

Calls for service are another distinct process. They can measure victimization, disorder, concern, nuisance complaints, duplicate reports, proactive officer activity, or differences in willingness and ability to call. A neighborhood with more calls may have more harm, more trust, more surveillance by residents, more conflict over public space, or some combination.

Arrests and stops are even more endogenous. They require police presence, observation, legal discretion, investigative priorities, and enforcement choices. For offenses such as drug possession, public drinking, loitering, traffic violations, and weapons possession, increased patrol directly increases opportunities to discover records that are then treated as evidence of future risk.

Feedback-loop causal diagram

Structural conditions and segregation
        │
        ├──────────────► Underlying victimization ◄──────────────┐
        │                         │                              │
        │                         ▼                              │
        │               Victim reporting behavior               │
        │                         │                              │
        │                         ▼                              │
        │                 Police-recorded incidents              │
        │                         │                              │
        │                         ▼                              │
        └──────────────► Forecast / risk model                   │
                                  │                              │
                                  ▼                              │
                         Patrol allocation and tactics           │
                           │              │               │      │
                           │              │               │      │
                           ▼              ▼               ▼      │
                     Deterrence or    Police-discovered   Stops, │
                     displacement     offenses/arrests    searches,
                           │              │               force
                           │              └──────► records ◄─────┘
                           │
                           ▼
                  Future underlying victimization

Patrol experience also affects trust and future reporting, creating
a second loop through the measurement process.

Ensign and colleagues mathematically demonstrated that repeatedly updating a model with discovered crime can create runaway allocation to the same neighborhoods even when true crime rates do not justify it. Resident-reported incidents attenuate the loop but do not necessarily eliminate it. citeturn26view5turn25search14

Lum and Isaac’s Oakland analysis supplied recorded drug events to a PredPol-like process and compared the resulting geography with survey-based estimates of drug use. The simulated patrol allocation concentrated in lower-income, minority neighborhoods and would have targeted Black residents at roughly twice the rate of white residents, despite survey evidence that drug use was much more broadly distributed. The lesson is not that every burglary forecast has the same disparity, but that police-generated labels can transform enforcement geography into apparent risk geography. citeturn2search0turn2search4turn16search7

Differential reporting creates a different failure mode. Akpinar, De-Arteaga, and Chouldechova showed that even systems trained on victim reports rather than arrests can shift hot spots away from high-victimization, low-reporting areas and toward medium- or high-victimization areas with higher reporting. The resulting allocation can simultaneously overpolice one area and underserve another. citeturn25search3turn25search11

Dirty data and historical police misconduct

Richardson, Schultz, and Crawford documented predictive-policing deployments in jurisdictions where source data had been produced during periods of unlawful, biased, corrupt, or otherwise unreliable police practice. Their “dirty data” framework is broader than ordinary statistical noise. It asks whether records embody unconstitutional stops, fabricated evidence, discriminatory enforcement, manipulated classifications, or institutional incentives that make them unsuitable for secondary algorithmic use. citeturn2search2turn2search6turn25search6

A predeployment audit must therefore identify:

\[ \text{record source} \rightarrow \text{legal authority} \rightarrow \text{collection practice} \rightarrow \text{known misconduct} \rightarrow \text{cleaning or exclusion decision}. \]

Deleting race from the model is not enough. Location is correlated with residential segregation, zoning, wealth, transit access, surveillance density, calls, and prior enforcement. Environmental features can operate as proxies for racialized urban policy without any explicit race field. Conversely, excluding all environmental variables does not eliminate bias if incident labels already reflect unequal enforcement or reporting.

Service patterns and repeat police presence

Police presence affects both the numerator and the meaning of crime rates. Consider two neighborhoods with equal underlying rates of low-level drug possession. If one receives twice the officer-hours, it offers twice the detection opportunity. Feeding resulting arrests into the model can justify still more patrol.

For victim-reported violent or property offenses, repeated presence may increase reporting by making police more accessible. A rise in recorded incidents after deployment could therefore mean increased crime, increased reporting, increased detection, or reduced trust accompanied by changes in which events are reported. A fall could mean deterrence, displacement, underreporting, victim withdrawal, or recording changes.

For this reason, evaluation should report separate outcome families:

Racial geography in actual deployments

The Markup’s broader Geolitica investigation found that forecast areas across 38 jurisdictions disproportionately covered neighborhoods with higher shares of low-income, Black, and Latino residents relative to the jurisdiction as a whole. That comparison does not alone establish unjustified discrimination because crime, population exposure, land use, and reporting also vary geographically. It does establish a distributive burden requiring explanation and causal measurement—especially when forecast-linked patrol may entail stops, searches, surveillance, and repeated presence. citeturn27view5

An adequate racial-geography audit should calculate, for each demographic group \(g\):

\[ \text{Forecast exposure}_g = \frac{\sum_{c,t} I(c \text{ forecast}) \times \text{population}_{gc}} {\sum_c \text{population}_{gc}}, \]

\[ \text{Patrol exposure}_g = \frac{\sum_{c,t} \text{officer-minutes}_{ct} \times \text{population}_{gc}} {\sum_c \text{population}_{gc}}, \]

and contact, stop, search, arrest, and force rates per resident, per ambient population, per reported victimization, and per officer-hour. No single denominator is sufficient. Resident population is weak for commercial districts; ambient population is difficult to estimate; reported crime inherits reporting bias; officer-hour denominators can normalize an already unequal allocation. A robust audit presents several denominators and explains their assumptions.

The relevant benefit distribution must also be measured. A heavily patrolled Black neighborhood may receive reduced victimization and faster response while bearing more stops and surveillance. Accountability requires estimating both rather than treating either enforcement intensity or crime reduction as the sole welfare measure.

Procurement, transparency, and minimum field-testing protocol

Procurement failures

Predictive-policing contracts have often been treated as ordinary software subscriptions even though the product affects government deployment of coercive authority. Conventional procurement questions—price, uptime, support, integration, and cybersecurity—are insufficient.

The Plainfield contract illustrates a basic value-for-money failure. The department paid $20,500 for the first year and $15,500 for an extension, generated tens of thousands of forecasts, but officials said the system was rarely or never used to direct patrol. A technically sophisticated product has no operational value if training, command policy, workflow integration, staffing, or officer acceptance prevents treatment delivery. citeturn26view0turn26view1

NYPD’s records illustrate a transparency failure. Vendor pilots were structured as free “gifts,” subject to nondisclosure agreements, and the department initially denied public-record requests by invoking nonroutine techniques, security, and proprietary interests. After litigation, records emerged piecemeal, but the in-house system reportedly did not store historical outputs. A free pilot can create switching costs, shape internal development, or transfer methods without appearing as a conventional procurement expenditure. citeturn24view1turn24view2

Acquisitions create another accountability gap. HunchLab’s transfer from Azavea to ShotSpotter meant that logs identifying variable importance in the Philadelphia models were unavailable to later evaluators. Geolitica’s closure and asset transfer likewise complicated the distinction among source code, patents, engineering knowledge, customer contracts, and successor functionality. Contracts that do not require escrow, documentation transfer, output retention, and post-termination audit access make public validation dependent on vendor continuity. citeturn26view0turn27view2

The most common deficiencies are:

• no publication of forecast denominators or historical outputs;
• no contractually specified baseline;
• “accuracy” warranties without a defined metric, area, offense, or horizon;
• vendor evaluation of its own product;
• nondisclosure provisions that obstruct public oversight;
• no requirement to preserve model versions and feature logs;
• no linkage between forecasts and patrol dosage;
• no audit of source-data legality and reliability;
• no disparate-impact or community-contact analysis;
• no termination trigger for nonuse or failure to beat baseline;
• no assignment of responsibility when a vendor is acquired or shuts down.

Contract requirements

A procurement should state that all forecasts, scores, model versions, input snapshots, configuration changes, officer-facing outputs, and dosage records are government records subject to retention and audit, with narrow redaction for genuinely sensitive operational details. Public evaluation can aggregate or delay release to avoid facilitating evasion; secrecy about active locations does not justify permanent secrecy about historical performance.

The contract should define a forecast event and specify whether multiple overlapping alerts count separately. It should state the eligible area, excluded land, temporal window, offense taxonomy, occurrence-time rules, geocoding procedure, late-report policy, and treatment of revised classifications. It should require forecast probabilities or comparable scores—not only colored boxes—so calibration can be tested.

Payment should be staged. Initial compensation may cover integration and a silent test. Continued payment should depend on data quality, operational use, independent baseline performance, and evaluation completion, not on unverified crime trends. Crime reduction itself should not be a simple performance bonus, because agencies and vendors could influence classifications, reporting, or forecast coverage.

Intellectual-property protection should not prevent:

• independent execution on a locked test set;
• inspection of feature definitions and transformations;
• subgroup and neighborhood performance testing;
• access to model cards and version histories;
• evaluation of calibration and false positives;
• public disclosure of aggregate metrics;
• preservation of evidence after contract termination.

Minimum field-testing protocol

The following protocol is the minimum defensible sequence before operational deployment.

Statistical design requirements

Power analysis must precede the trial. Microplace crime is sparse and overdispersed; a study with very few districts or events can miss policy-relevant effects. The unit of randomization should minimize spillover while providing enough units for inference. If commanders cannot prevent officers from crossing nearby grid boundaries, randomizing larger clusters or time blocks may be preferable.

The primary analysis should be intention-to-treat:

\[ Y_{it} = \alpha_i + \gamma_t + \beta Z_{it} + \epsilon_{it}, \]

where \(Z_{it}\) is randomized assignment. A secondary treatment-on-treated estimate can use assignment as an instrument for actual dosage if exclusion and monotonicity assumptions are plausible.

Count outcomes should generally use Poisson or negative-binomial models with exposure offsets and cluster-robust or randomization-based uncertainty. Rare violent outcomes may require longer periods, pooled prespecified categories, or Bayesian partial pooling, but post hoc category expansion should not rescue a failed primary endpoint.

Forecast and causal endpoints must be separately preregistered. A model can win the forecast endpoint but fail the crime endpoint. A patrol intervention can win the crime endpoint even when the proprietary model does not beat the baseline. These are substantively different findings.

Public reporting template

Every published evaluation should include:

\[ \begin{array}{ll} \textbf{Data:} & \text{source, report type, exclusions, missingness, geocoding error};\\ \textbf{Forecast:} & \text{unit, horizon, lead time, coverage, probability or rank};\\ \textbf{Baseline:} & \text{method, tuning rules, analyst information};\\ \textbf{Accuracy:} & \text{precision, recall, PAI, calibration, uncertainty};\\ \textbf{Operations:} & \text{officer-minutes, tactics, compliance, contamination};\\ \textbf{Causality:} & \text{randomization or identification strategy};\\ \textbf{Spillovers:} & \text{displacement and diffusion};\\ \textbf{Equity:} & \text{forecast, patrol, contact, and benefit distribution};\\ \textbf{Cost:} & \text{license, integration, analyst, officer, evaluation, exit};\\ \textbf{Reproducibility:} & \text{code, data schema, outputs, model versions};\\ \textbf{Governance:} & \text{authorization, complaints, audit, sunset, termination}. \end{array} \]

Public release need not disclose active patrol boxes, current staffing vulnerabilities, or real-time response rules. Historical outputs can be delayed, spatially generalized where necessary, and evaluated inside secure environments. Tactical sensitivity is compatible with meaningful algorithmic accountability.

Final assessment

What adds value beyond conventional hot-spot policing?

Conventional hot-spot identification plus focused, procedurally constrained patrol has the strongest evidence base. The aggregate literature indicates modest crime reductions and does not support the assumption that crime simply moves next door. Visible patrol, problem-oriented responses, and environmental changes can work when delivered with sufficient dosage. citeturn3search2turn3search6turn22search11

Near-repeat forecasting adds plausible short-term information for selected high-volume offenses, especially burglary and vehicle crime, where recent events can temporarily elevate nearby risk. Its incremental value is likely greatest where event times are timely and reliable, near-repeat chains are common, patrol can respond quickly, and the baseline is not already a high-quality dynamic hot-spot map. Its value is likely small where reporting is delayed, event volume is sparse, or chronic locations dominate.

Self-exciting point-process models have demonstrated forecast gains in important trials, particularly the Los Angeles/Kent work. The reported gains over analysts and short-window count maps should not be dismissed. However, the evidence base is too narrow and too connected to PredPol’s developers to establish a general, jurisdiction-independent advantage. Independent multicity replication against tuned transparent baselines remains the missing test. citeturn18view2turn27view0

HunchLab showed that richer features can be operationalized, but the strongest field evidence points more clearly to patrol strategy than to proprietary model superiority. In Philadelphia, a dedicated marked car reduced property crime, while lower-dosage or less-visible conditions did not. Since HunchLab was not randomized against simple recent-count hot spots, the study cannot determine whether gradient boosting, demographic and environmental variables, or mission optimization selected materially better places. citeturn18view3turn27view2

Risk-terrain modeling offers the most distinctive conceptual contribution when it changes environments rather than merely intensifying patrol. It can identify modifiable features and support lighting, property management, business practices, code enforcement, or place redesign. Its forecasting literature suggests meaningful spatial ranking ability. Its causal crime-reduction claims remain weaker when interventions are bundled and evaluated through before-and-after trends. RTM should be judged most favorably as a transparent diagnostic framework linked to noncoercive remediation, and more skeptically when it functions mainly as another rationale for concentrated stops and enforcement. citeturn22search1turn22search28turn22search31

PredPol/Geolitica’s public evidence is mixed and ultimately insufficient to justify broad vendor claims. A developer-connected experiment reported meaningful performance and patrol effects. An independent Plainfield audit found extremely low alert precision, low incident recall, and almost no operational use. Santa Cruz stopped using the system and later banned predictive policing. LAPD’s inspector general could not isolate program effects from available records. Geolitica’s corporate shutdown and asset transfer further complicated accountability. citeturn23search7turn23search13turn26view0turn27view0

Shreveport provides the clearest caution against attributing value to a map without controlling operations. Its randomized design was undermined by low power and heterogeneous implementation, and it found no significant incremental reduction. Its central recommendation remains the right standard: compare predictive and traditional maps while holding interventions and effort constant. citeturn21view0turn21view1

Police-built tools are not presumptively more accountable than vendor products. They may avoid subscription costs and trade-secret barriers, but undocumented code, absent version control, inaccessible outputs, and internal-only validation can make them equally opaque. NYPD’s inability to recreate historical outputs is an example. citeturn24view1

Overall evidentiary rating

Final judgment

The technically defensible conclusion is narrower than either vendor promotion or categorical rejection.

Place-based crime concentration is real enough to support focused prevention. Transparent recent-event counts, historical persistence, kernel-density maps, and analyst knowledge already capture much of that signal. Near-repeat and self-exciting models can improve short-term ranking for some offenses, and environmental models can identify stable risk settings. Yet the public evidence does not show that proprietary predictive-policing products reliably and generally produce crime reductions beyond well-implemented conventional hot-spot policing.

Where positive effects have appeared, they are often inseparable from visible patrol dosage, problem-oriented intervention, or environmental remediation. Where products have failed, the causes include poor precision, weak recall, nonuse, missing dosage, incompatible workflows, low event volume, and inadequate evaluation. The relevant procurement question is therefore not “Does the algorithm predict crime?” It is:

\[ \text{Does this system, compared with a transparent alternative, produce enough additional public benefit to justify its financial cost, enforcement burden, data risks, and loss of accountability?} \]

On the evidence available through August 2, 2026:

• A jurisdiction can justify conventional microplace hot-spot policing when it uses lawful, proportionate, evidence-based tactics and monitors displacement and community harm.
• Near-repeat or self-exciting forecasts may merit a controlled trial for timely, high-volume property offenses, but only after they beat tuned recent-count baselines in a prospective silent test.
• RTM may add the most public value when it redirects policy toward modifiable environmental conditions rather than repeated enforcement.
• HunchLab-style mission optimization may improve operational discipline, but its forecasting and routing components must be evaluated separately from the effect of dedicated patrol.
• PredPol/Geolitica did not establish a sufficiently independent, reproducible, multicity evidence base before its corporate exit.
• Rebranding predictive functions as “precision policing,” “resource routing,” or “mission management” does not remove them from audit.
• Acoustic gunshot detection, crime mapping, real-time crime centers, and investigative pattern matching should not be called predictive policing unless they actually estimate future risk and influence prospective allocation.
• No system should be deployed without preserved outputs, complete dosage measurement, independent baseline comparison, randomized or credible quasi-experimental evaluation, racial-geography analysis, public governance, and an enforceable sunset clause.

The minimum responsible default is therefore transparent hot-spot analysis first, proprietary prediction only after demonstrated incremental value, and coercive patrol only after separate evidence that the chosen intervention reduces harm more than it creates.