The Trouble with Banks' Risk Models: Q&A with the Chief of SAS
One of the latest data discovery products, announced Thursday by SAS Institute, promises to let non-IT folks in banks do complex big data modeling from their own computers.March 22
James Goodnight is a big-time executive who knows "big data" and all the buzzwords that preceded it and, he says, so do bankers.March 15
Jim Goodnight, who co-founded Cary, N.C., analytics software company SAS Institute in 1976 and has been its CEO ever since, came to New York Wednesday to tell journalists about the company's high-performance and visual analytics software. Goodnight's conversation with Bank Technology News veered toward the efficacy of risk models.
BTN: What's the state of the art in analytics right now?
Goodnight: High performance analytics. That's a very big push right now in banking. If you can make the modelers much more productive, running 10 to 100 more models to make the very best investment they can find, instead of running 4-5 times in a week, that can make a big difference in banks.
BTN: Wall Street firms use high-performance computing to run pricing and risk models. Are you seeing this used a lot in more traditional banks?
In retail banking, it's all about making sure you're gaining new customers. The emphasis is on customer intelligence and marketing, to know who you should be marketing to. We're seeing a lot of that.
BTN: When you say "high-performance computing," are you talking about in-memory computing [in which calculations and queries are performed on data stored in local memory, rather than on a separate storage device]?
Yes. It's not only in-memory computing, but also using hundreds of processors rather than a single processor. Almost everything we do today is one processor beating itself to death trying to compute billions of operations. We've found it's much better on some larger problems to enlist hundreds of processors. They've gotten so darn cheap, for $10,000 you can buy a blade with 256 gigs of memory and 32 processors running in parallel, so you ought to take advantage of the hardware. Companies like us need to suck it up and get it done. That's what we've been doing for the past three years, trying to get all of our software running in parallel so that we can solve much bigger problems.
BTN: Have you accomplished that?
Yes. It's all done, we'll come out with another major release in June or July of high-performance analytics.
BTN: What has adoption of high-performance analytics been like in banking?
It's been very high. Right now, we have quite a few situations where people are taking another look at risk. After the London Whale did its job over there, everybody's concerned about risk. We have high-performance risk analytics where we can do 100,000 market simulations in a matter of 15-20 minutes, where it used to take 15-16 hours.
BTN: What do you think of the state of banks' risk models? The Fed recently conducted stress tests of the largest U.S. banks. The banks' own stress tests had much rosier results than the Fed's.
The only difference would be the pricing algorithms, because if you define the factors that are to be stressed, it's just a mere computation of what that market value will be, plus or minus 20% of that risk factor. Only the pricing models would be different.
BTN: Do you think some of the pricing models need to be tweaked?
Every bank thinks they have the best pricing models in the world. I'm not sure they'd tweak them.
BTN: There was a lot of debate after the financial crisis about how good banks' risk models are, a lot of risk managers never thought housing prices would drop. They made adjustments later, in hindsight.
In risk simulation, you're simulating the state the market has been in the last two years. If it makes a dramatic move that hasn't occurred in two years, your risk model is not going to be accurate.
BTN: So it's the assumptions plugged into the models rather than the models themselves that were at fault.
Since you're simulating market states around the average of the past couple of years, the mean is still the same for all the risk factors. If you do 100,000 simulations, you're still going to have the same mean. That bothers me because in fact, in the Black Swan events, that mean is going to change. It gives you, at the upper end of the distribution, some idea of how much danger you're in.
BTN: So you still need people around who have some sense of change coming.
If a lot of structured investment vehicles had been properly priced, the risk would have shown up, but we were pricing these housing bonds as if they were AAA government bonds. The risk is not going to show up. There were a lot of people at fault who were not pricing right.
BTN: What do you think the next big bubble will be?
Probably housing again.
BTN: Can it happen again?
Prices are going up rapidly in our area and we're building all over Wake county, it's the number-one growth county in the country. A lot of houses are going up, the prices of existing houses are rising. There's always been a housing bubble every 10 years, hasn't there?
BTN: I don't think like the last one.
Not like the last one. The last one was exasperating because they were taking credit default swaps on all this stuff. They figured they were insured, they didn't think about, what if the insurance company goes under. That was the main reason AIG was bailed out, to keep the banks from going under.
BTN: Is there anything that you take away when you look at what happened over the past five years, to say risk modeling should change in a certain way?
During the housing boom, some companies sent people out in the field to talk to someone who was selling a house, they found out there was no money down, the borrower was borrowing 110% of the value, on and on. That's how bad loan originations had gotten during the housing bubble. Several banks realized that because they had people in the field checking out this stuff. You can't rely on numbers all the time, you need verification in the field.
BTN: What will be the main driver for the use of analytics in the next five years?
Customer analytics, including fraud. If you pull credit and debit card data, DDA data and loan information together, you could have a single view of the customer that will let you do fraud, marketing automation, and customer intelligence at the same time. We're working with several banks right now to build that single view of the customer that lets us run all the bank's models on that one set of data.
BTN: That's been a Holy Grail for banks for some time.
It has been, but there's been a lot of fighting on it. We're up here talking to people in compliance, their job is to keep the CEO out of jail, they don't give a darn about marketing. The data ought to be shared with the marketing people because 90% of it is the same. Yet both departments are paying to do the extraction from the databases. At some point, the CEO needs to step forward and say all these silos need to start working together to get a single view of the customers. It has to come from the CEO, because everybody wants to protect their silo, that's their job and the job of the people who work for them.
BTN: What's the state of the art in mining customer data? There's so much data out there, you can look at people's Facebook posts and Twitter streams, you can buy data about how they pay their phone bills, their rent, etc.
Yes. It turns out that a lot of that external information you're paying for can be derived from the data you've already collected. We were working with one bank on credit card data, our modelers could tell us when somebody was getting a divorce, or if someone's gotten married or had a baby, you can tell just by looking at the names of the stores they shop at.
BTN: What's on your product drawing board?
This year we'll have two more releases of visual analytics this year, one in June and one in December. By December we should have every feature anyone's ever asked us for.
BTN: What are some of the new features?
Some of the new ones are scenario playing and forecasting, where you can say, what if you take into account this or that variable? What would happen if I could get 10 more reps in field, how will that affect my revenues? All of that is available on an iPad.