Headlines about used car loan companies imploding may not get much attention at insurance companies, but this “old news” from the Fed should: Those companies have been making risk assessments based on wonky data and are paying the price. (The Effects of Credit Score Migration on Subprime Auto Loan and Credit Card Delinquencies).
There is a punch line here for auto insurers. You are relying on the same wonky data, so your risk scores will likely perform worse than historically expected, too.
False negatives
A false negative occurs when a fact you intend to observe is not visible. A classic in the literature is a pregnant woman who takes a pregnancy test that returns a negative result. In that case, you can blame the test. But a more subtle case is what the Fed is showing now, where the problem is with the data.
Data is missing that typically weighs down a credit score and is thus driving scores higher—while the riskiness remains the same.
Imagine a historically stable data process where good and bad observed data drive positive and negative features that calibrate a risk score. If a time or place existed where bad things were not tracked as usual (thanks, COVID) or penalties were simply less enforced (thanks, COVID), then risk scores would rise for no good reason. The COVID timeframe encouraged a period of financial transaction forbearance unlike any we have experienced in modern times.
Auto insurance has other false negatives, too. Having less enforcement of traffic rules (thanks, COVID) and less availability of traffic courts (thanks, COVID) caused similar problems with reporting on motor vehicles. For example, running red lights may have produced no tickets -- still very risky behavior, just with no typical negative indication on record. The same with speeding tickets, which haven’t been issued as frequently in recent years.
The simple equation of score = intercept + good factors - bad factors means that the absence of a bad factor mathematically leaves a score in better shape than it deserves.
Decades of observable data have been used to establish that a credit-based risk score can be useful in describing a risk scenario where the higher the score the lower the risk in an auto insurance relationship. We tend to pull credit data on assessing a new risk in policy acquisition and on a routine basis when we re-score entire portfolios of policies.
But the sort of financial forbearance that is showing up in bad loans for used cars means that people have had virtual clemency. This let many lower-quality risks appear with higher credit-based risk scores, so more risk decisions were made at terms and conditions unwarranted by the true riskiness.
Another risk – and a new source of data
Credit data is observed backward but applied forward. A forward-looking data stream that has recently been introduced can complement existing credit-based risk assessment methods (JSI CDPI).
The stream, which grew out of work to assess the risks of students seeking loans, tracks risks based on macroeconomic effects on occupational categories. As cars were taking off a century ago, being a buggy whip maker put you at risk. Restaurant workers had a rough time during COVID. Generative AI is currently a threat to clerical workers. And so on.
Incorporating indices linked to wages and wage opportunities can help adjust the false negatives in current credit-based scores.
For all the insurance decisions linked to credit-based scores, this may be a learning moment.
