Inside Portland police’s new data-driven approach to catching stolen cars

Waving a deck of cards, former software developer Michael Terrett gave a group of officers in Portland Police Bureau’s east precinct a crash course in probability.

He dealt cards around a conference room. Officers seated came from across the Portland metro area. They’d signed up to help the Portland police catch stolen vehicles that November night. They came from nearby police departments and sheriff’s offices.

Before they could hit the road, though, they had to learn Portland’s new tactics: looking at cars on the road like a poker player looks at cards in their hands.

Terrett dealt cards around a conference room, and asked officers that night if they had watched poker on TV and seen how a widget on the screen calculates in real time a players’ chances of winning.

When it comes to finding stolen cars, he said, “We can do the same thing.”

For years, the bureau catalogued only the bare details of stolen vehicles: make, model, year and color. But in 2021, Portland police were floundering under a record-breaking wave of auto thefts.

Terrett, who left his tech job and joined the bureau in 2018, convinced his bosses to start tracking vehicle information with his former industry’s insatiable appetite for data. They could calculate not just how many Toyotas were stolen in any given month, but how many had broken windows or tinted windows.

“It really is all about probabilities and statistics,” said Dr. Jeff Tyner, of OHSU’s Knight Cancer Research Institute. Tyner and his team helped Terrett and PPB develop its first datasets.

The result seems to be a Silicon Valley-esque disruption of old police tactics. It’s not unlike Spotify using your past music taste to suggest a new artist. Or, in Terrett’s example, like playing poker.

“The goal is that we stop less vehicles, but our outcomes are better,” Terrett told the officers.

Auto thefts peaked in 2022, according to bureau data. Portlanders had reported 10,903 stolen vehicles, breaking a nearly three-decade-old city record for auto thefts in a calendar year.

This year numbers have plummeted, with 4,204 reports to date.

In the conference room, Terrett put the bureau’s trove of data into action.

He flicked a widget onto a screen, measuring the likelihood that a car was stolen. Then he told the officers that each card in front of them would represent a detail about a hypothetical vehicle and driver.

For example: Does the car have a broken window? Or dealer license plates? Has it been spray painted? Maybe all of the above?

“If there are 10 cars out on the road and one of them has a stack of (factors), there’s a higher probability that that vehicle is stolen,” Terrett said.

Terrett asked a Clackamas County deputy to flip over his top card. It read “Toyota pickup.” Terrett’s widget flickered to 4%. That, Terrett explained, showed the likelihood the vehicle was stolen based only on its make and style.

“All we know is that it’s a car on the road,” Terrett told the room. “What’s your next card say?”

The deputy flipped over his next card: “Missing plates.” The widget ticked up to 17%.

“Not bad. I would like better, but not bad,” Terrett said. “So, you stay with the car and find something else. What’s the next thing you find?”

Over the next few minutes, the widget rose and fell based on every card the deputy flipped: factory-installed tinted windows, erratic driving behavior, a bunk trip permit on the back window.

The deputy ended up with a 33% chance that the car was stolen, based on bureau data.

Terrett walked through the same exercise with other officers’ cards. Some clocked lower odds, some higher. One reached 66%.

“So if we have a set of cars, this is the one we certainly want to focus on,” Terrett said.

Game theory

That night, Gresham police officer Jack Labuhn was among more than a dozen officers zigzagging near the Portland-Gresham line on a search for stolen cars.

Around 7 p.m., he and his partner spotted a lifted Ford pickup truck. It had no license plates – cause enough to be pulled over.

By looks alone, Labuhn acknowledged, the truck didn’t seem like a sure hit. He had participated in 13 stolen vehicle missions. Data would show, for example, a likelier suspect to be a Hyundai Sonata with a spare tire and a broken rear passenger window.

But Portland police’s training also factors in how drivers behave when a patrol car pulls up near them. As they maneuvered their patrol car closer, the truck’s drivers abruptly banked into a gas station.

“That’s a diesel,” Labuhn said. The truck had pulled into a gas station without a diesel pump.

The police stayed on the road and eyed the truck as they tried to pass nonchalantly. Labuhn’s eyes met the driver’s. Then the truck turned around and tore out of the parking lot.

Reinforcements descended on the scene. Other officers threw a spike-strip in front of the truck.

The truck then bailed down a quiet residential road. Two occupants hopped out and fled into an apartment complex. They were ultimately arrested.

Circling the truck, Terrett catalogued factors for the bureau’s dataset. The ignition had been cut out for a multitool to jury-rig the starter. Clothes and junk – evidence of being “lived in,” Terrett said – piled in the backseat. A burn hole in the door handle suggested someone had speared it with a hot metal object to break the lock. They found needles and gun ammo.

Labuhn, who said this was an ordinary stop, started his policing career two years ago and uses the probability method regularly.

“I think every single agency, to be frank, in the country could do the exact same thing,” Labuhn said. “A lot of this is due to training.”

The data-driven approach wasn’t available when Terrett joined the force. He said he often found himself on the night shift with a veteran officer who confidently pointed out stolen vehicles as they passed.

It was insight based on experience, a gut feeling. Terrett asked the officer how he knew and the officer shrugged off the question as if unable to fully explain. Seasoned officers had basically written their own databases with each stop, year after year.

“It would have been years of repetitions,” Terrett said.

With data collection, they can prove it beyond a hunch and weigh which factors are more suggestive than others.

Tracking factors like broken windows, tint, and spare tires may seem obvious now, but Terrett said that’s the point. The database distills the bureau’s collective knowledge into something easily parsed by newer officers, like Labuhn.

“You can take an officer who is new and get them to be really precise at identifying stolen vehicles in a very short period of time,” Terrett said.

Still, as any poker player would say, there’s no such thing as a sure thing. Terrett pulled behind a black Audi without license plates that night, for which Terrett could pull him over.

And it had self-installed window tint, a positive indicator. Terrett wanted to see if they would start driving erratically.

“We have probable cause to stop the car but we’ll see if we get any driving behavior changes,” Terrett said.

The car kept driving smoothly. Terrett still pulled him over, only to discover the driver was an ex-convict who had legally purchased the car. The driver didn’t have a valid driver’s license, so Terrett waited for the driver’s mother to come pick it up.

“We could tow this. You cannot be driving this,” he told the driver. It was a warning. “You got to get it registered. Get it a title put into your name, as well.”

After the stop, Terrett again entered the data, explaining it was maybe more important to enter the data on vehicles that aren’t stolen as it is on the ones that are.

“Because we need to know what not to look for,” Terrett said. “The more data we analyze, I think, the more precise and better we can be.”

On average, according to bureau data, police used to stop more than 30 cars before finding one stolen.

The night of the mission, police stopped 16 cars. Ten were stolen.

Numbers versus randomness

Portland law enforcement officials have many reasons to feel confident about the strategy.

Auto thefts in the city have taken such a significant statistical drop that Terrett and his colleagues have drawn attention across the country. They recently spoke with Interpol, the international police organization, and presented at a national public safety conference held by the U.S. Department of Justice.

They’ve also collected headlines. Portland police have arrested some prolific car thieves, including two men last October who were linked to more than a hundred thefts of Teslas in Oregon and Washington.

Cody Linderholm, an assistant district attorney who has prosecuted several car thieves, said police are more surgical and are limiting stops of innocent Portlanders.

“What I’ve seen in my work is it’s a small group of bad actors doing a lot of the heavy lifting,” Linderholm said.

Spoiling car thefts lead to other criminal disruptions, too. Police often find evidence of other crimes, whether it’s illegal drugs or weapons, evidence of burglary, or identity theft.

“It’s not just somebody steals a vehicle and then they’re captured in the vehicle, right?” Linderholm said. “It’s somebody captured inside a stolen vehicle and then there is evidence of other crimes – which may or may not be charged depending on the circumstances and evidence.”

But factors outside their control could affect the bureau’s success at any moment – and how police, criminals and lawmakers react.

For example, Kia and Hyundai-brand cars were more vulnerable just months ago when car thieves went viral for turning ignitions with USB cords. Thefts vaulted across the country until the manufacturers issued a recall.

Some car thieves have also been reprogramming keyfobs to steal American-made cars, Linderholm noted.

“Oftentimes, especially in a community like Portland, it’s fluid. It’s always changing,” Linderholm said.

Case law could also change. Four years ago, police in Marion County arrested a man for sitting in the driver’s seat of a parked, stolen Honda Accord. They arrested him at gunpoint. Last month, the Oregon Court of Appeals overturned his conviction, saying police lacked probable cause.

It’s a Darwinian struggle that cancer researchers, like those who helped Portland police develop their dataset, are familiar with.

“In some cases there can be adaptations that occur that allow (cancer) cells to evade therapy and then we observe drug resistance,” said Tyner, the OHSU researcher. “We can observe very similar principles at play with the stolen vehicle operations where selective pressures are applied, perhaps words get out about it and things change and behaviors adapt.”

Many types of crimes – not just auto thefts – rose and fell nationwide after the pandemic. Tyner said that they have not yet analyzed how much of the drop in auto thefts could be attributed to the pandemic.

Yet tracking data will forever be useful for quickly spotting which direction factors are trending, Tyner said.

Even if auto thefts were to tick up again, the fact that police are finding more stolen cars with fewer stops proves the data’s effectiveness to Terrett. The police are smarter, the community is less impacted.

“To get this level of precision takes a lot of discipline and patience,” Terrett said. “I may be out here for 10 hours and I might make one stop – but that one stop I nailed it.”