PredPol's crime predictions came true less than one time in 200

The software borrowed its math from earthquakes

George Mohler, an applied mathematician, and Jeff Brantingham, a UCLA anthropologist, built their model on a different hazard first: the equations seismologists use for earthquake aftershocks. In that math, a location carries a background rate of risk, and each event nearby temporarily raises the risk around it before that decays away.¹ Applied to burglary instead of tremors, one burglary raises the modeled risk on its block and its neighbors before it falls back. Tested against 5,376 burglaries the Los Angeles Police Department had logged in the San Fernando Valley in 2004 and 2005, the earthquake math beat a standard hotspot map. At the same 10 percent of the city flagged, it caught 660 of that year's 2,627 burglaries, against 547 for the hotspot method.¹

Starting in November 2011, patrol officers in Los Angeles's Foothill division received a shift map with 20 target boxes, each 500 feet by 500 feet. It was built from nothing but the type, place, and time of crimes already reported: no arrest records, no demographic data.² Mohler and Brantingham incorporated the software as PredPol in 2012; the company renamed itself Geolitica in 2021, and by 2019 roughly five dozen U.S. police departments paid for its forecasts.³

PredPol scored places, not people, unlike Chicago's heat list, built to rank individuals, or the risk scores judges use at sentencing, which score a person only after arrest. It differs, too, from ShotSpotter's gunshot sensors, which flag a sound that happened rather than forecast one.

Plainfield's boxes, checked against what happened next

In February 2018, Plainfield, New Jersey's police department began paying Geolitica $20,500 a year for its forecasts.⁴ Over the next ten months the software issued 23,760 predictions, each a box, a shift, and a crime type.⁵ In 2023, The Markup and Gizmodo checked each one against Plainfield's own crime reports for a match, excluding 129 where an officer had a recorded visit. That left 23,631 predictions to check against reality.⁵

The match rate for robbery and aggravated-assault predictions was 0.6 percent: 32 hits out of 5,722. For burglary it was 0.1 percent: 8 hits out of 10,141.⁵ A more forgiving measure, crimes preceded by a matching prediction, still found only 10 percent.⁵ The reporters' own summary: "The Markup's analysis suggests that Geolitica predicted crime at a rate slightly better than blindly throwing darts at a map."⁵

Plainfield's police captain, David Guarino, said the department had rarely, if ever, used the software to direct a patrol, and had already let the contract lapse. Geolitica did not respond to The Markup's methodology or findings, before or after publication.⁴ In 2019, a separate audit at the department where the software was born reached a similar dead end. Of 24,247 recorded visits, 74 percent lasted under a minute and 56 percent lasted 30 seconds or less, brief enough that officers appeared to be "driving through or past the designated location" rather than working it. The inspector general could not determine whether the software had reduced crime at all.⁶

Los Angeles discontinued its own program in April 2020, citing pandemic budget cuts, though the coalition that had organized against the software for years credited its own pressure instead.⁷ Santa Cruz, an early pilot city, banned government use of predictive policing outright in 2020, requiring any future system to be shown "scientifically validated and peer reviewed" and free of bias first.⁸

Recorded crime is not crime

The reason a hit rate this low can coexist with real bias starts with what the software was trained on: never crime itself, but crime that got reported, and crime officers found by being sent somewhere. Those are not the same; most crime is never reported at all, and The Markup's researchers cite federal survey data finding fewer than half of violent-crime victims report to police.⁵ That gap creates a loop: patrols go to a box, officers record more of whatever crime does get reported, the model reads the rise as confirmation, and the next shift's map favors the same box again.⁹¹⁰

Kristian Lum and William Isaac tested the loop directly in Oakland, California. Estimated drug use, from a national health-survey rate applied to the city and unrelated to any policing decision, came out roughly even across racial groups; recorded drug arrests clustered hard in a few lower-income, mostly Black and Latino neighborhoods.⁹ Feeding the arrest data, not the survey estimate, into a PredPol-style model and simulating a deployment targeted Black residents at roughly twice the rate of white residents, and feeding each round's "discovered" arrests back into the next round's training made the skew worse each time.⁹

A 2018 paper made the loop mathematically exact. Ensign and his coauthors wrote out PredPol's own equation and modeled two neighborhoods with nearly identical true crime rates, 10 percent against 11. A system that updates only on crime its own officers discover, they proved, converges with certainty to sending patrols to just one of the two, and running actual PredPol code on Lum and Isaac's data reproduced exactly that collapse. As the authors put it, PredPol "critically conflates amount of crime and incident data": it predicts where incidents will be recorded, not where crime occurs.¹⁰

The other half of the explanation this desk has already taught in full: a system built to catch a rare event will be wrong most of the time even when it works as designed. The event almost never happens in any one place on any one shift. That is how a hit rate under one percent and a real racial skew can both be true: one comes from how rare the event is, the other from whose past enforcement the forecast was trained on.

The question the company's best evidence never answered

The standard defense of a number this low is that a patrol car sitting in a flagged box can itself prevent the crime it predicted, so a simple count scores the empty box as a miss. Geolitica never made that argument about The Markup's findings, since it did not respond to them. But Plainfield's numbers weigh against it anyway: officers had a recorded visit at only 129 of the boxes, and folding those back in changed nothing.⁵

The company handled an earlier, different complaint differently. In December 2021, The Markup and Gizmodo published a nationwide bias analysis, finding predictions concentrated far more heavily in Black, Latino, and low-income neighborhoods than white, higher-income ones.¹¹ Then-CEO Brian MacDonald responded, calling the dataset "incomplete" and "erroneous" and naming a specific error, without disputing the predictions were genuine. He confirmed the founders' 2018 simulation had found the same pattern up to 400 percent stronger, and that a bias-corrected version tested less accurate. Adopting the fix, he said, "would reduce the protection provided to vulnerable neighborhoods with the highest victimization rates."¹¹

The founders had already published a narrower rebuttal using real arrests: their 2018 field-trial paper found no statistically significant racial difference in arrests between algorithm- and analyst-directed patrol days, three LAPD divisions, not the nationwide deployment the 2021 analysis covered.² Mathematician Andrea Bertozzi, a field-trial co-author and, by her own disclosure, an investor in the company but not a founder, offered a broader defense a year earlier: "there is no merit to the accusations." She cited peer review and the trial's drop in crime, without addressing the accuracy critique or the feedback-loop papers in print.¹²

The company's strongest evidence predates the accuracy exposé by eight years: the 2015 field trial reported algorithm-directed patrols cut crime 7.4 percent, about twice the analyst-directed reduction.² The paper sits behind a paywall; its figures are reported as two independent readers state them, a follow-up by two founders and the Los Angeles inspector general's 2019 audit, citing the same figures.²⁶ Neither substitutes for the original study, but both had reason to check it first.

A record of decisions can pass for a forecast

Geolitica shut down at the end of 2023; SoundThinking absorbed its staff and intellectual property but not the crime-prediction code itself.⁴³ The 500-foot boxes are gone; the trap did not need them. Any system trained to predict where enforcement should go, from a record of where it already went, can produce the same closed loop Ensign proved: training data cannot tell the true rate of a thing apart from the historical intensity of looking for it in one place.¹⁰ That holds whether the record is arrests, citations, or audits.

Santa Cruz's 2020 ban put a version of the right question into municipal law: any future system has to be shown "scientifically validated and peer reviewed" before a council can approve it again.⁸ The earthquake math Mohler and Brantingham started with can meet that test, because the ground it forecasts does not respond to being watched.¹ A model trained on where officers were sent is training on human decisions, and predicting from those decisions mostly returns them, relabeled as tomorrow's forecast.