To catch a thief

A young mathematician at SCU has helped equip police in Santa Cruz and L.A. with an algorithm that predicts where crimes might happen next. Is this the future of policing?

To catch a thief
Data driven: When it comes to predictive policing, Officer John Shepard of the Santa Cruz P.D. is a believer. In fact, it's one of the reasons he joined the force. Photo by Charles Barry

Sergeant Frank Albarran, a tall, muscular, 16-year veteran of the Los Angeles Police Department, gazes out the windshield of his black and white cruiser and shakes his head with a mixture of disbelief and disdain.

We’re parked on a quiet residential street in North Hollywood, all single-family homes, shade trees, and tidy lawns. It’s the middle of a beautiful, sunny day. There’s no one and nothing in sight that looks remotely suspicious. In fact, there isn’t anyone in sight at all. “I don’t know why we’re here,” Albarran mutters.

Albarran has been dispatched to this unlikely spot because of the work of a young assistant professor of mathematics based some 350 miles away at Santa Clara University. George Mohler is a pale, bashful 30-year-old who happens to be helping to mastermind one of the most talked-about innovations in modern American crime fighting. Along with several other scholars at the University of California, Los Angeles, he is developing one of the most promising experiments in an emerging field known as “predictive policing.” The idea: Although no one can know for sure when an individual might commit a crime, it is possible to forecast patterns of where and when homes are likely to be burgled or cars stolen by analyzing truckloads of past crime reports and other data with sophisticated computer algorithms. The algorithm Mohler and his colleagues have developed is influencing the work of hundreds of police officers across Los Angeles and in Santa Cruz—and yielding impressive results.

“We rank areas according to risk,” says Mohler. Identifying which areas are likeliest to see a rip-off can help police figure out where to deploy officers to prevent the crimes from happening in the first place. But why the algorithm flags a particular area—like this quiet North Hollywood block—as risky, Mohler acknowledges, is not always “intuitively clear.”


Mohler got his undergraduate degree in mathematics (and his slap-shot training) in his native Indiana, and went on to the University of California, Santa Barbara, where he researched mathematical modeling of polymers and fluids. After graduating, he got a job offer in that field but found himself more intrigued with a stranger one. Two UCLA professors, anthropologist Jeffrey Brantingham and mathematician Andrea Bertozzi, were working with the LAPD to develop algorithms to predict crime. They saw Mohler’s résumé and wanted him on board; turns out some of the mathematical models Mohler had been working with that describe pattern formation in polymers were similar to those the UCLA professors were using to predict burglaries. “I read their papers, and it made a lot of sense,” says Mohler. “I thought what they were doing was really cool.” He took the job.Mohler isn’t exactly in close touch with the mean streets. He spends most of his days in a sparsely furnished office in the basement of O’Connor Hall on the Mission Campus, at the end of a subterranean corridor bedecked with posters advertising upcoming math conferences and job openings for computer scientists. He’s on the skinny side of thin, with an easy, oversize smile that gives him an almost alarmingly cheerful look. With his rectangular geek-chic glasses, lace-less Converse sneakers, plaid shirt, and flop of black hair, he could be a San Francisco website designer or an indie rocker—which he has been, actually. In his spare time, he played bass in a couple of bands. The music has gone on hiatus since he and wife Courtney Elkin Mohler, an assistant professor in SCU’s Department of Theatre and Dance, became parents in 2011. But George Mohler still manages to strap on skates for some ice time as part of an adult hockey league.

The team gathered years’ worth of data from the LAPD on the time and place where home and car burglaries and auto thefts had taken place. (They focus on those crimes mainly because there are lots of them, providing a rich data set.) One of their key early insights was that crime tends to beget crime: If a house gets broken into, the probability of neighboring houses getting broken into soon after rises. Most crimes, like burglaries and car thefts, are not planned in advance but are opportunistic: A bad guy sees an unlocked window and ducks in. “Burglars typically don’t travel far. They tend to commit crimes in their own neighborhoods,” explains Mohler. “They have a lot of information: They know when their neighbors are at work and which houses are easy to get into. And when they succeed, they go back again. You see it in the data.” Mapping those patterns can give police an edge in figuring out where to deploy extra cars and cops to catch bad guys—or, better yet, keep them from opening that unlocked window in the first place.

In some ways, the notion of predicting where crimes will happen based on where they’ve happened in the past is obvious. That one event increases the likelihood of similar events occurring nearby in space and time is well established in other fields of research. In fact, you can see it everywhere in ordinary life: A punch thrown in a bar increases the chances of more punches. One kiss leads to another. Analysts and academics use the principle more methodically to predict where banana trees might be found, or where corporate defaults will cluster. One of Mohler’s main contributions to work on a new model for predictive policing was to find and adapt an algorithm developed by seismologists to help predict where aftershocks will strike after an earthquake.

If there have been a lot of muggings on a particular street for the last 50 weeks, there will probably be some the following week. Cops know that, of course. But the idea is to make those assumptions and guesses more accurate and to turn up patterns that aren’t so readily apparent.

Corporations have long used similar predictive analytics to anticipate consumer demand, and have found that cross-pollinating data can yield unexpected results. A famous example comes from Wal-Mart’s analysis of what its customers in coastal areas stock up on before hurricanes. The list includes duct tape and bottled water, naturally, but also a surprise item: strawberry Pop-Tarts.

Analyzing crime data can similarly yield counterintuitive conclusions. Most people think good lighting makes an area safer, for instance, but studies have found that it actually increases the chances of being victimized. It seems that muggers want to be able to see their potential targets clearly.

Hot spots: Not just where crime has been, but where it's likely to occur next.
Hot spots: Not just where crime has been, but where it's likely to occur next.
Future tense: Mathematician George Mohler focuses on geography—and what (and where) crimes may come. Photo by Charles Barry
Future tense: Mathematician George Mohler focuses on geography—and what (and where) crimes may come. Photo by Charles Barry


In August 2010, the team’s work, though still in the theoretical stage, spawned an article in the Los Angeles Times. That caught the attention of Zach Friend, a crime analyst with the Santa Cruz Police Department. “I called up Mohler, who had just taken his job at Santa Clara,” says Friend, 33. “I said, ‘We’ll take this out of the classroom and put it into the field, if you’re willing.’” The team agreed.

Friend brought Mohler in to help sell the idea to his colleagues. The cops met the mathematician with a certain amount of bemusement. “Our nerd pal!” one officer calls Mohler when I visit the department’s headquarters one spring day. “He’s bringing corduroy back!” yuks another. Still, the brass bought in. “We’ve had budget cuts like everyone else. Resources are scarce, and we need to use them as efficiently as possible,” says Santa Cruz Police Chief Kevin Vogel. “I thought it would be worth giving this a try.”

So every workday for the past two years, Friend has come in early to type the time and geocoded location of the most recent burglary and auto theft reports into the department’s computer system. Mohler’s algorithm then crunches those reports together with the last seven years’ worth of crime data, and spits out a map of Santa Cruz with 10 boxes on it, each representing an area 500 feet long and 500 feet wide—about half a block. Those are the hot spots that the algorithm deems likeliest to see thefts that day. The maps are handed out to officers at the beginning of each shift. They cruise through the boxes when they have time in between active calls.

“We’re very pleased with the results,” says Chief Vogel. In 2012, burglaries were down about 7 percent compared to 2011.

And the program has drawn international attention. Time magazine, NPR, The New York Times, and news crews from as far away as France and Germany have reported on it, and scores of other police agencies, as well as the Department of Defense, have gotten in touch.


One of the most interested out-of-town cops was LAPD Captain Sean Malinowski, an athletically built 46-year-old with hair receding from a sun-reddened brow. Malinowski helped coin the term “predictive policing” in an influential paper he co-authored in 2008 with then-Los Angeles Police Chief William Bratton.

Malinowski worked for several years as executive officer to Bratton, who is a near-legend in American law-enforcement circles as the police chief on whose watch crime plummeted first in New York, then in L.A. He was always getting invited to give talks on the future of policing, and part of Malinowski’s job was to brainstorm with him about what to say. One of Bratton’s key innovations was a computerized system called CompStat, which collects detailed reports on crimes and maps where they were committed. Versions of the system are now used by police across America. Thinking about ways to build on and improve CompStat’s data-driven approach, they came up with the idea (and catchy title) of “predictive policing” and wrote about the concept for the Oxford Journal of Policing.

With Bratton’s towering reputation behind it, the idea caught fire. Soon after the article’s publication, the National Institute of Justice organized a conference on predictive policing, and the Department of Justice handed out more than $1 million in seed grants to a fistful of police departments interested in pursuing the idea.

Office Shepard on the job. Photo by Charles Barry

Various agencies are now trying out different approaches, pulling in all kinds of data. In Arlington, Texas, cops have created maps overlaying residential burglaries with building code violations. They found that as physical decay goes up, so do burglaries. They’re using those findings to deploy police more efficiently. In Tennessee, University of Memphis criminologists and local police are using business-analytics software to compile crime reports and layer in variables like weather, lighting conditions, and proximity to concert venues, along with reporting from PDA-equipped beat cops, to find connections. The system noticed, for instance, that colleges’ spring-break week reliably spawns a rash of burglaries. And in Minneapolis, a special Crime Analysis Unit identifies locations where gun crimes have been reported, then factors in geographic details on things like bus routes and proximity to parks, liquor stores, and public libraries. Combining that with seasonal data enables them to predict when certain public parks and other areas will become arenas for gun violence.

The LAPD, meanwhile, has continued working with the team that includes Brantingham at UCLA and Mohler at SCU. Impressed with the results their algorithm seemed to be getting in Santa Cruz, Malinowski got approval to put it into practice in Los Angeles starting late in 2011.

Los Angeles means a trial on a completely different scale. Santa Cruz is a famously laid-back town of only 58,000 people; in 2012 it saw a total of two homicides. Los Angeles is a sprawling metropolis of more than 3 million, where someone gets killed almost every day.

Faced with the size of the city and its police force, Malinowski has been introducing predictive policing one division at a time. The North Hollywood division, which patrols a chunk of the San Fernando Valley that is home to some 204,000 people, was the second, beginning to use Mohler’s algorithm early in spring 2012.

“Auto theft, burglary from vehicles, and residential burglaries are down 16 percent compared to the same period last year,” Captain Justin Eisenberg tells me inside the North Hollywood division headquarters, a sprawling modernist building on a busy, sun-blasted avenue. “That’s pretty incredible. Predictive policing isn’t a panacea, but it is surprisingly useful.”

The program is now running in a handful of divisions, with trials under way in each of the city’s 21 divisions. Crime has dropped in the divisions where the program is already established. Neither Malinowski nor the scholars at Santa Clara and UCLA are ready to say that all that crime reduction is due to the algorithm. “But everyone sees we’re getting great numbers and wants it in their area,” Malinowski admits.


Like the earthquake version of the algorithm to predict where an aftershock will strike, mathematician George Mohler’s “expectation maximization” algorithm for predictive policing models the incidence of two sets of phenomena. The first is the “background rate” of events that occur randomly—earthquakes or spontaneous crimes. Different geographical areas have different background rates: Some areas have more geologic faults, and some neighborhoods have social and economic factors that subject them to more crime.

The second is the “branching structure” of similar events generated by the random, spontaneous ones. With earthquakes, that’s aftershocks. With burglaries, it’s more burglaries, which often beget still more burglaries, spreading out like a tree’s branches. You can further narrow the list of likely locations of future crimes by ruling out areas where they are impossible; you can’t have a residential burglary in the middle of a park, for instance.

Mohler and his team analyzed several years of data on the time, place, and type of property crimes in Los Angeles to see the patterns of where and when they occurred and were followed by others. (Murder and other violent crimes are far less frequent and so don’t offer enough data for the algorithm to provide a reasonably accurate forecast, says Mohler.) From there, they derived a set of mathematical functions to predict both the occurrence of spontaneous crimes and the probabilistic distribution of where the crime “aftershocks” are likely to occur. The result: a literal road map to future crimes. VB


The idea of predictive policing has its critics. Civil libertarians are concerned it could result in extra police pressure on poor and minority neighborhoods. If a cop spots someone holding a bag and looking at a building on a street the algorithm has flagged as a likely spot for burglaries, he may be more likely to stop and frisk the loiterer, points out Andrew Guthrie Ferguson, an assistant law professor at the University of the District of Columbia, in a recent paper. The officer might catch a thief—or might open himself up to a charge of racial profiling.

There are also some unnervingly Minority Report-ish law-enforcement experiments under way that use predictive techniques to help determine whether an individual is likely to commit a crime. Pennsylvania probation and parole officials are working with a University of Pennsylvania statistician who has developed an algorithm to help estimate the risk of specific inmates re-offending after release. And the Department of Homeland Security is experimenting with a system of sensors that tracks airport passengers’ heart rate and other physical indicators to help determine which should be singled out for an extra search. Mohler stresses that his team’s algorithm looks only at geographic areas, not individuals. “We don’t put demographics into the model,” he says. “There’s no individual information being fed in.


On a more practical level, hard-headed street cops are understandably skeptical about the whole notion. Back in Albarran’s cruiser, we move on to another box indicated by the algorithm, a block on busy Ventura Boulevard. “This box here,” says Albarran, jabbing a finger at the map, “it doesn’t tell us what crime or who to watch out for. We know this is a busy street with a lot of stuff getting stolen out of parked cars. We don’t need predictive policing to tell us that.

“I personally don’t think it’s very helpful,” he grumbles. “Most of my guys feel the same way.” The whole point, though, is not for Albarran to spot a crime—it’s for his presence to stop one from happening.

“We don’t see this as a way to arrest people but to deter crime,” says Mohler. “The LAPD is probably among the most sophisticated departments in the United States,” he goes on. “They do a good job, but we’ve shown you can do better.” Lt. Albarran might scoff at that. But police around the country are paying attention. Mohler and his colleagues launched a commercial version of their program, dubbed PredPol, in 2012, and have so far made sales to several police departments and fielded inquiries from scores more. As mobile technology develops, so will the program—and the interface, as well as what kind of information is provided to officers. For predictive policing, it seems, the future is looking good.

Make AI the Best of Us

What we get out of artificial intelligence depends on the humanity we put into it.

The Co-Op

Santa Clara University has long been a bastion of interdisciplinary learning. A new fund is taking cross-collaboration to new heights.

Human at Heart

How Santa Clara University is distinguishing itself as a leader in one of the fastest-growing industries in the nation.

A Campus on the Rise

New buildings on campus—count ’em, six in total—aren’t the only changes brought by a successful $1 billion fundraising campaign. Come explore what’s new.