A survival model of how small business loans end, built on every loan the SBA has guaranteed since 1991.
A loan leaves the books one of two ways, paid off or charged off, and the two compete. Both depend on how long the loan has been outstanding, on what was lent to whom, and on the economy it lives through. The model estimates all of that at once on 1.37 million loans, and the lab projects a book five years forward under the Federal Reserve's 2026 scenarios. Built with Next.js and TypeScript; the model is estimated by a committed Node script from SBA, FRED, and Federal Reserve sources.
35.9%
had been charged off within ten years. For the loans approved in fiscal 2013, the figure is 5.6%. Everything between those two numbers is what the model has to explain: what was lent, to whom, and into which economy. It puts a book like fiscal 2025's originations at 3.0% charged off within five years on the Fed's baseline path and 4.4% on the severely adverse one.
Open the labConsumer credit is managed by vintage. Every book I have worked near gets watched as a set of curves: loans booked in a given quarter, tracked by how many months they have been outstanding, compared against the quarters before them. When a curve bends the wrong way early, that is the first sign something changed in the underwriting or the economy. I have read hundreds of those curves. I wanted to build the model that produces them.
The SBA's 7(a) file is the one loan-level credit dataset in the public domain that carries an outcome and a date for every loan: 1.96 million approvals since fiscal 1991, each with its term, size, program, sector, and the day it paid off or the day it was charged off. That is a different kind of data from the aggregate series behind Nine Quarters, and it asks for a different kind of model.
The first blind test I ran came back with zero payoffs before 2005, which is not a result, it is a symptom. Every loan paid off before a system migration in the middle of that year carries a payoff date of the last day of May or the last day of July 2005: 281,010 loans, a fifth of all the payoffs in the file. Charge-off dates survive back to 1991. A survival model that took the file at its word would learn that loans from the 1990s never paid off young and then all paid off at once.
So the equations are estimated on loan-quarters from 2005 Q4 onward, with older loans entering the risk set at whatever age they had reached by then, the way a study handles subjects who enrol late. Charge-off curves by vintage are still valid for every year, because with every loan resolved the cumulative charge-off rate is simply charge-offs over loans, whenever the payoffs were dated. Payoff curves are shown only from fiscal 2006, and the blind run starts from the last quarter before the crisis instead of 2004.
Each quarter a loan is outstanding it does one of three things: it continues, it pays off, or it is charged off. The model is a pair of hazard equations, one for each exit, estimated together against continuing on 25.01 million loan-quarters from 1.37 million loans. Standard errors are clustered by loan, because a loan contributes a row for every quarter it survives.
Time on the books enters as 19 steps, and that curve is the point. A new loan almost never charges off in its first two quarters. From there the hazard climbs steeply: the odds in the third year are 6.1 times the odds at the end of the first, and they stay near that level for years before easing past the tenth. Lenders call the climb seasoning. Term matters more than anything else in the file: a loan of three years or less carries 2.5 times the charge-off odds of a four-to-seven-year loan, and a loan over twenty years, which in this program means real estate, carries 0.08 times. A business under two years old carries 1.26 times the odds of an established one.
The economy enters where it acts. Each point of unemployment in the borrower's state raises the charge-off hazard 6% today, and each point it rose over the year before last adds 8%: a business that failed while unemployment was climbing is charged off about a year later, and the file records the charge-off. I tested lagged levels and changes at every step from zero to two years, and that pair won. Payoffs move the other way. Each point of unemployment moves the payoff hazard down 7%, because a business is harder to sell or refinance in a weak economy, and a one-point rise in the prime rate over the past year moves it up 5%, since most of these loans float over prime and a reset up is when a borrower who can refinance does.
log odds of exit e in quarter t = αe(age) + βe · loan + γe · economyt, for e ∈ {paid off, charged off} against continuing.
118 parameters, 1,365,446 loans in the estimation window, 183,055 charge-offs and 885,963 payoffs. Newton-Raphson, converged in 11 steps.
83 of every 100 loans that leave this file leave by paying off. The textbook survival curve treats those as if they had simply stopped being observed, and the loans that pay off are the loans that were never going to default. Treat payoffs as censoring and the fiscal 2013 cohort reads 13.0% charged off within ten years. Count both exits properly, the way Aalen and Johansen showed in 1978, and it is 5.6%.
That is not a rounding error, and it is the reason the model has two equations. A charge-off hazard on its own cannot say what share of a book will eventually be lost, because the answer depends on how quickly the rest of the book leaves through the other door.
I re-estimated the model on the quarters before the crisis, 2005 Q4 through 2007 Q4, and ran the fiscal 2005 to 2008 cohorts forward on the unemployment and prime rate path that actually followed. It puts fiscal 2007 at 10.8% charged off within four years, against 24.0% in fact, and running it again with the economy frozen at each loan's booking quarter barely moves it. That is the finding, not a failure of the test: in nine quarters of a strong economy the model had nothing to learn about unemployment, and its coefficient came out at -0.017 per point. A model estimated before a recession does not know what a recession does.
So the decomposition uses the full model, whose sensitivity was learned across the crisis. On the realized path it puts fiscal 2007 at 13.7%; with the economy frozen, 8.5%. The recession explains 5.2 of the 15.5 points between a calm economy and what happened, and the rest is what the file cannot see: how those loans were underwritten. A second test fits through 2015 Q4 and scores every loan-quarter from 2016 Q1 to 2019 Q4: AUC 0.76, predicted charge-offs 0.634% a quarter against 0.407% observed. With the crisis in its sample the model expected more losses than the long expansion delivered, which is the direction a reviewer can live with and is disclosed rather than tuned away. The lab shows it decile by decile.
| Fiscal year | Actual | Pre-crisis model | Full model | Economy frozen |
|---|---|---|---|---|
| 200584,302 loans | 9.5% | 8.7% | 10.3% | 9.6% |
| 200686,028 loans | 16.2% | 10.1% | 12.2% | 9.0% |
| 200788,028 loans | 24.0% | 10.8% | 13.7% | 8.5% |
| 200861,470 loans | 19.1% | 11.2% | 12.6% | 8.7% |
Share charged off within sixteen quarters. The pre-crisis model is fit through 2007 Q4; the full model runs on the realized unemployment and prime rate path, then with both held at each loan's booking quarter.
The charge-off date is administrative. SBA books the loss after the guarantee is purchased and the collateral is liquidated, often a year or more after the business closed, so the hazard here is the hazard of a loss being recognized. The file has no delinquency history, no payment history, and an interest rate for only half the loans, which is why none of those are in the model. Collateral lowers the charge-off odds by 21% in the window the model is fit on, and it carried the opposite sign when I fit the whole file with its broken payoff dates. A coefficient on something the lender chooses measures the lender's judgment as much as the loan, and it should be read that way. Loss given charge-off is the gross balance SBA recorded, 73% of the original approval on average, before recoveries and before the split between the lender and the guarantee. The uncertainty bands cover the coefficients and nothing else; the scenario paths are taken as given.
Since April 2026 the SBA has published the delinquency status of open loans, which would let a production version model the pipeline from thirty days past due to charge-off directly instead of reading it through the age curve. That is the first thing I would add. The second is a lender's own book in place of the industry file, with the macro sensitivity estimated here carried in as a prior. And a second team: the backtests here are the developer's own, which is the situation model risk management exists to fix.
Data: SBA 7(a) FOIA file as of 2026-06-30, FRED, and the Federal Reserve Board. Model build 2026-09-30.