Machine learning is helping auto lenders make better decisions

https://www.bloomberg.com/opinion/articles/2019-01-05/algorithms-are-determining-your-creditworthiness

by resalisbury • 8 years ago

37 45 8 years ago

benj111• 8 years ago

The title has been changed from the article, and I think slightly misses the point.

Plus ML can be more biased than humans (or more biased than we will admit to) https://www.bbc.co.uk/news/technology-45809919

ralph842• 8 years ago

I guess it's, like, the author's style to use the ..., like, ... construction, but that makes it no less annoying.

csa2• 8 years ago

After a quick scan, I only saw this used once. Does it really bother you that much?

That said, the ellipsis construction is used in an attempt to make text feel more like casual conversation — in this case, something like drawing out the preceding word in a questioning manner or a maybe a pregnant pause. This is a very standard device used in the register that Matt Levine uses for his column, and I imagine that most people find that these casual devices he uses makes his column more accessible for a wider audience.

ralph842• 8 years ago

Three times. There's nothing wrong with writing informally for a wide audience, but in this context it just comes across as "How do you do, fellow kids" rather than making the content any more accessible.

dhh2106• 8 years ago

Here's the main study cited in the article: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3301009

"We find that the same minority applicants who faced lower approval rates on auto loans do not face lower approval rates on credit cards, even during the same year. This finding suggests that the human element of auto lending, rather than differences in hard information, leads to the lower approval rates for minorities."

mhuffman6• 8 years ago

This insinuates that car lenders are racist.

I am truly curious of what the business model of denying minorities a car loan is?

I mean, long ago, I guess I can at least process the ugly mindset of "we don't want 'them' as our neighbor", but I can not come up with any logical or illogical reasonable reason why a business person would intentionally lower their income just to be racist.

Literally no one would know that you made a loan to a minority, so that can't be it.

Is the insinuation that the loan officer might "get a feeling" that a person might not repay a loan based on the race?

If that were the case, at least in the financial sector, it would seem that another financing institution would have noticed this inefficiency long before 2019 and gobbled up all of that free money on the table.

I also wonder if the higher interest rates of credit cards accounts for higher approval rates.

crooked-v• 8 years ago

I am truly curious of what the business model of denying minorities a car loan is?

Why does it have to be a business model? Humans are not rational actors. The segregation era almost certainly lost tremendous amounts of money for retail businesses as a whole, but their management still went along with it. The remaining implicit bias in many businesses still loses tremendous amounts of money for retail businesses as a whole (see the various "wealthy black celebrity treated like a shoplifter" news stories, for example), but people still do it.

jlarocco• 8 years ago

If that were the case, at least in the financial sector, it would seem that another financing institution would have noticed this inefficiency long before 2019 and gobbled up all of that free money on the table.

Of course. They'll just only hire people who aren't subconsciously biased. No problem. They'll just put "No racists please," in their job advertisements. Simple as that. LOL!

michaelchisari• 8 years ago

You're assuming rationality.

skywhopper2• 8 years ago

You seem to assume racism is always an active choice, rather than largely (these days) a set of systemically biased policies as well as implicit and often unconscious prejudice. It's also not all about rejecting loans. Rather, it's that the terms offered are different for every buyer, and the groups that are being discriminated against are charged higher rates, and thus pay more, for the same product.

There's plenty of data out there proving that, for whatever reason, African American loan applicants with identical credit scores and income are routinely charged higher interest rates for auto and home loans than their white counterparts. The next question to ask is "why?" but we can't pretend the effect doesn't exist.

lkdjjdjjjdskjd2• 8 years ago

How do auto loans and credit card approvals differ? Surely both have some human element to it still?

Maybe the sums involved are different, so giving loans is more risky than approving credit cards?

Also, incentives might be different. Maybe credit card companies want "bad" borrowers, because they make the most money if people go into high debts on credit cards? On the other hand for normal loans there is no such incentive, because the interest rates are fixed and much lower.

clairity• 8 years ago

most people get auto loans st the dealership with a salesperson (the human element), but get credit cards via online or paper forms (no human involved).

bluedino• 8 years ago

Car loans are secured (the bank owns the car). If you buy stuff on a credit card, you own the stuff

sologoub2• 8 years ago

This is where things get complex with ML. Any dimension that is correlated with the protected categories can potentially get you into a biased and discriminatory outcome. Because embedding are not necessarily transparent, it’s hard to evaluate what the model attaches to.

The most accessible example is anything to do with fine grained location, such as zip+4. Because people tend to cluster, zip+4 has been identified as a close proxy for race and other protected classes.

When we are talking ML-derived decisions and the nuanced approach, we have to be cognizant of the fact that certain repayment patterns and behaviors could also be strongly correlated to a protected class.

darawk2• 8 years ago

What if the correct predictor really is the zip code, though? Let's say there's a city that's predominantly black, and it has a statistically significantly higher default rate than other cities. However, the default rate in that city is high for residents of all races who live there. In this scenario, the 'true' predictor really is the city, but it also correlated with race. Is simple correlation to a protected class sufficient to nix a variable?

skywhopper2• 8 years ago

I'm not seeing any evidence that the ML approach doesn't have the same or similar biases. I'm sure the results are different but we can't say they're better just because a computer generates them. In fact, we've seen many times that ML may often tend to give implicit biases even more emphasis, because in an attempt to create "objective" measures against which to judge a dataset, ML creators inject their own biases, or rely on data which is itself biased in its collection or categorization. (So, for example, if zip code is a major indicator in the ultimate ML algorithm, it may be completely "objective", while still reinforcing biases against minority applicants, simply because of how wildly self-segregated many US cities are, or because of self-reinforcing effects of systemic bias in other, human-controlled parts of the economy.)

hakfoo• 8 years ago

Geography may be a factor.

You might reasonably apply by mail for a credit card from an out-of-state bank, but a lot of auto lending revolves around a purchase made in a nearby dealer.

If the dealers in minority-dominant communities tend to be more aggressive about lending qualifications, that could explain it all neatly.

Though, my blind guess would be that you'd see more "buy-here/pay-here" and "no-credit-required-but-the-loan-is-at-34-percent-interest" dealers in minority neighbourhoods (skewing poorer), so approval rates would likely be at par or higher. Or are they qualifying "approval rates" with "at sane market prices?"

apocalypstyx2• 8 years ago

"Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them."

" 'Thou shalt not make a machine in the likeness of a man's mind,' " Paul quoted.

"Right out of the Butlerian Jihad and the Orange Catholic Bible," she said. "But what the O.C. Bible should've said is: 'Thou shalt not make a machine to counterfeit a human mind.'

sgt101• 8 years ago

Oddly, prescient.

lifeisstillgood• 8 years ago

The Dune trilogy / quadrilogy whatever yes?

apocalypstyx2• 8 years ago

It's actually a hexalogy, but, yes.

mr_toad• 8 years ago

The prescription against thinking machines, along with similar plot devices (shields promoting hand-to-hand combat, the defacto ban on atomic weapons) allowed FH to construct a quasi-medieval SF setting.

I wouldn’t take mystical pronouncements in fictional settings intended to advance a plot too seriously.

m0zg2• 8 years ago

Makes sense. Just don't use race, gender, or ethnicity as a feature in the model and those factors _per_se_ won't be a consideration. Their correlates might still be a consideration, however, so it's not an entirely "fair" system, in the social-justice sense. Bias generated in meatspace will still affect it indirectly.

anewguy9000• 8 years ago

wouldnt a system that is "blind" in the sense that it has no prejudice be free to use those (anything) as features? isnt the whole point of the NN to get loans repaid?

m0zg2• 8 years ago

Yes, but then you end up in a morally fraught territory where someone can claim that you ended up rejecting the loan because someone is a woman or gay or whatever. If those factors _directly_ aren't a consideration, your position is a lot more defensible.

Story time: Google ad serving system used to think (and probably still does) that if someone is interested in engineering topics per their browsing patterns, they're a male, much to the consternation of female Google engineers.

If gender wasn't a feature in the system at all (an unrealistic proposition in an ad targeting system), this problem would not arise as there would be no way for a system to determine if someone is male or female due to the absence of the corresponding feature, and therefore descriptive vocabulary. More defensible, at least, if not "fair".

Iwan-Zotow2• 8 years ago

"make better decisions" as making themselves richer and consumers poorer?

dan-robertson• 8 years ago

If you make bad decisions by offering loans at too low a rate, you will lose money on average and eventually go bankrupt.

If you make bad decisions by offering loans at too high a rate because you have high overheads or you’re cautious or your actuaries are wrong or you’re racist then people won’t take out your loans because they get a better deal elsewhere or the loan is too expensive so you lose money by not profiting from making the loan.

Now there’s a balance here: you want to be competitively priced so people will take out your loans but you also want higher margins so you can make more money for investors or expand or make less money in other places to attract customers.

I think in this case “make better decisions” means either morally better (ie less racist), or better in the sense of “make better deals to win more customers to make more money”

Iwan-Zotow2• 8 years ago

What kind of function ML is optimizing? Who specified that function? Lenders made their decision from the dawn of time, so we have some data to compare. Now (say, last year) they made decisions with help of ML - what changed? Who got a better deal?

Machine learning is helping auto lenders make better decisions

Related Stories