Can ‘predictive policing’ software help city police prevent crime?

A growing number of police departments are using software programs that companies claim can predict where crime is going to happen.

When police officers decide which blocks to patrol, they usually look at where crime has happened in the past, and they rely on their own hunches. But a growing number of city police departments are now considering something else: the calculations of software programs designed to predict where crimes will happen in the future.

That’s the case in the city of Reading, where every morning during roll call, police officers are handed maps — printed off a computer — of their patrol districts. A few spots on each map are highlighted with red boxes. Those boxes are supposed to show the areas at highest risk for certain crimes during that shift.

The predictions are spit out by a software program called PredPol, created by a company with the same name. The Reading Police Department launched the software in 2013 for about $18,000 a year. Dozens of other police departments in the U.S. and abroad have purchased the same program.

PredPol runs off of some pretty basic data: the type, location and time of previous crimes. The software plugs that data into an algorithm that’s based on math used to predict earthquake aftershocks. The reason is that scientists say after an earthquake, you’re more likely to see aftershocks nearby. The same is thought to be true of certain crimes.

Applying a mathematical model from physics to crime is not as strange as it might sound, says John Hollywood, a senior scientist at the RAND Corporation who studies predictive policing. “It’s very common to take approaches from one scientific area and then adapt [them] to another scientific area,” Hollywood said.

But even if the math is reasonable, how does PredPol work in practice?

Life in the red boxes

One morning after getting his map, Reading police officer Jacob Stefani heads to his patrol car. He tucks the map into the car’s visor and pulls out of the station.

Stefani says he spends a lot of time driving around the red boxes, trying to catch criminals in action, or even deter them from committing a crime.

The police department has set the software to predict property crimes.

“We’re having a lot of daytime burglaries right now,” Stefani says. “So you’re looking for maybe some young kids skipping school, breaking into houses. Looking for anything out of the ordinary. Open doors, broken windows.”

At one point, the officer sees a man kicking the door of a breezeway. He gets out and asks the guy for ID. Turns out, he lives in the house and locked his keys inside.

Stefani also looks for people squatting in some of the neighborhood’s abandoned houses. He spots one house that’s boarded up with plywood. It looks like someone tried to pry the wood off the door.

Stefani radios in to the dispatcher, and then gets out of the car to investigate.

Stefani walks down the breezeway, slowly making his way to the backyard. He quickly gives the all clear. There’s no one in the yard, but it’s littered with empty soda cans and the screen door is broken. There’s gang graffiti on the wall. 

Stefani says it’s good to have guidance from the software. But on days when he finds nothing in the red boxes, it’s hard to tell whether the algorithm is working. 

“If no crime is happening, you don’t really know if that’s the result of you being [a] deterrence or just because that particular criminal didn’t really feel like conducting a crime.”

Does it work?

As you might expect, PredPol says its software works. A randomized controlled trial study published in the Journal of the American Statistical Association in 2015, by the researchers who created the software, showed PredPol was between 1.4 and 2.2 times as effective as crime analyst predictions in Los Angeles and in a town in England.

If the software weren’t working, it’d be hard to know. PredPol’s contract with the city of Reading requires the police department to participate in press conferences and joint marketing on behalf of the company, something its CEO says is a reasonable obligation. A 2013 article in SF Weekly brought attention to the same terms in PredPol’s contracts with other cities.

Different crime prediction software applications are being tested by Philadelphia and Atlantic City, New Jersey, among other cities. These programs use different algorithms, and there’s little independent research on their accuracy.

But one thing these software programs have in common is they’re harnessing computers to process data criminologists say is important. Human beings could do the same thing, but it would take a lot of time, says PredPol CEO Larry Samuels.

“The advantage of computers is they just process on and on,” Samuels said. “They’re able to process huge amounts of information, and they’re able to do it at incredible speeds.”

A 2013 report from the RAND Corporation said predictive policing algorithms are not a crystal ball. And small cities that don’t have a ton of crime data don’t need to shell out for crime prediction software, the report says. There are free tools they could use instead. In fact, the application Atlantic City is using is free. 

Reading Police Department officials say the software is working for the city, citing a recent drop in crime; the annual number of burglaries in the city decreased by about 18 percent from 2013 to 2014 and is on track to be lower this year.

But the number of burglaries in Reading has gone up and down for a decade. And Police Chief William Heim says introducing other crime-fighting measures like security cameras may have also contributed to a drop in crime.

“I don’t think any one tool is going to yield the results we’re seeing,” Heim said. “But the combination of them I think is really good.”

At least, until he does get that crystal ball.

Updated 11/4/15: We’ve updated this story to remove the implication that the software used in Atlantic City, New Jersey was created by a for-profit company.