Agencies just getting started on artificial intelligence and predictive analytics can face internal inertia. In that case, starting small may be helpful.
That’s according to Michael Collins, a statistician in the Federal Student Aid Office at the Education Department. Speaking during the Agency Readiness for Predictive Analytics panel Tuesday, presented by the Advanced Technology Academic Research Center, Collins described how his office, which administers the FAFSA form, uses predictive analytics to forecast loan default.
Now that federal student loan payment forgiveness, collection halts and waived interest – due to COVID-19 – is set to end Dec. 31, FSA has begun targeting its communications at borrowers they think are most vulnerable to defaulting on their loans. But the agency does this predictive modeling at multiple stages of the loan lifecycle as well.
“We’ll predict that when an applicant first files their FAFSA form, and applies for aid, and we’ll do it again at a point in time in which they actually are in college, and they’re getting their aid disbursed. And then we’ll also model it again, when they stopped going to school, either have graduated or dropped out, and model a probability of default at that time,” Collins said. “One of the most powerful predictors of whether or not a borrower will default on their loan is whether or not they finished the program and graduated. And we won’t know that until they either finished or not, until they’ve gone to repayment.”
FSA can use that probability of default to forecast future cash flows, plan outreach and message testing. It’s possible borrowers are unaware of repayment resources and options, he added.
“We’ll have some borrowers who just forgot, like maybe they have a six-month grace period after they finish school to start the payment. And they just didn’t do that. And that’s clearly going to be a much different situation than someone had been making payments for five years, and then suddenly stops making payments,” Collins said. “So someone starts getting late on their payments, and they haven’t defaulted yet – we want to help them stay out of default.”
Predictive analytics can also help detect internal failures.
Chakib Chraibi, chief data scientist at the National Technical Information Service at the Commerce Department, said his agency uses predictive analytics in collaboration with organizations to optimize resources. It helps to mitigate risk by proactively moving resources to areas where something could likely go wrong, he said.
Most commonly, predictive analytics is used in forecasting, whether with regression modeling, classification modeling or other types.
Working with the Department of Health and Human Services’ Office of Inspector General, NTIS has been using artificial intelligence and advanced analytics to spot suspicious transactions and stop improper payments or fraudulent schemes. By detecting anomalies and overpayments, Chraibi recalled one anecdote of a nail salon that charged Medicare as a surgical operation.
“Looking at patterns – and that’s a difficult area in predictive analytics. So you have to do it step-by-step,” he said. “First, you have to collect the data that you need. And then you need to develop a system about how to detect anomalies and that includes subject matter experts that are there.”
Data visualizations are helpful for spotting patterns that are “out of whack.” Data collection can also include public data from social media, news publications and other sources.
And sometimes NTIS has to work with agencies such as the Labor Department, where a lot of data is in stovepiped applications making it difficult to do effective predictive analytics, Chraibi said.
“We are working on establishing a data infrastructure that will help them integrate all their data and help each unit to actually get a better understanding of what the data tells them – not just within their unit, but across the agency,” he said. “So we have also made headways in processing errors, in claim processing, authenticating and verifying the eligibility of claimants, and so on and so forth.”