Hacker News Reader

Comment by basementcat | Hacker News Reader

Is there a market opportunity for completely automatically generated security ratings? Individuals in most developed countries are now assigned a completely computer generated "credit score" (from payment history and applications for credit); is it much more difficult to algorithmically assign an analogous score for a bond or similar instrument?

EDIT: Or is the problem a misalignment of incentives? It appears that investment rating providers are paid by the managers/sellers of the securities rather than the prospective buyer of the securities. This may create an incentive to attract more business by inflating ratings. If prospective investors paid for the ratings, perhaps there would be a different incentive alignment (and also possible opportunities for tragedy of the commons).

View on HN

I was a former ABS banker and I personally think there is an opportunity to automate the structured finance ratings process.

Here is how it works today. An originator finds people to lend to, aggregates them, works with a bank to structure them into a security that gets rated and then institutional investors buy the securities. Rating agencies get paid to apply their ratings criteria which they publish, allowing you to reverse engineer the model.

The opportunity, in my opinion, is a point of sale system where you hand iPads out to car dealerships, clinics whatever. Someone wants to make the big purchase, puts their social in and gets funded by the institutional investors on the spot. An institutional investor at the moment won’t get a say on funding, the platform will. They will simply set their high level criteria and the system will continue funding to get to some average that fits it.

For example, I could be an investor that says I want $10mm of subprime auto loans per month with average FICO if 615 and min/max loan sizes of XYZ. That gets entered into the system and then the dealerships continue handing iPads at the point of sale. The cash flow of the loans can move through the waterfall and ratings can be real-time in the flow as opposed to the security.

Anyway not sure if this makes sense as it is nuanced and I am being high level.

Real time cash flow seems like overkill,props if you can do it though. You see some lenders in the merchant cash advance market doing similar POS flow analysis. Lots of structured products have liquidity facilities to bridge short term issues so credit quality throughout the cycle tends to matter more than ST flow imo.

The problem is that this doesn't track correlations between people. The implication is that each user has an independent risk, but that's not always true. In particular it was the problem with CDOs during the last crisis: people assumed that one person's default would not cause another's, and what seemed like a collection of moderate-risk loans was actually very high risk in aggregate.

If you're aware of that fact then you know what you're getting and can deal it. But it's the tacit assumptions, magnified by aggregation and automation, that caused a disaster.

I think the big issue with that is those institutions don't want to be acting as loan servicers. Someone needs to do all the work of processing payments, working to define alternate payment plans in some cases and doing foreclosures/reposessions when someone defaults. Who does that in your model?

1. When you say 'credit score', I presume you mean the score assigned by the credit bureau. This isn't an objective measure, and sophisticated lenders develop their own scores (using raw+summary data from the bureau and other sources).

2. Analysing the ability for a large/complex company to pay back a bond is pretty complicated, as it depends on understanding the operations of the company, and the riskiness of the cashflows generated by those operations.

3. Your point about misalignment of incentives is correct, and mentioned in the article.

The problem with that idea is that as soon as you define objective criteria, people will find a way to game those objective criteria.

how is this different from the current rating system the article is about?

The answer to your question is almost certainly no. This is because credit ratings are not based on payment history, or anything as simple as that. Instead, credit analysts have to understand how what can be complicated legal structures work.

Can legal agreements be expressed as an algorithm?

If they could in all cases then we wouldn't need courts.

Not to replace courts but to understand legal agreements better. If you could express it as an algorithm, perhaps, you could rewrite in simpler english or answer questions about it?

When logically consistent, but the written contract is not always the final word. Example: lack of consideration (https://www.nolo.com/legal-encyclopedia/consideration-every-...)

You'd still run into the halting problem. And you'd have lawyers wanting to write legal agreements that run into that, so that nobody can determine what they really mean, so that nobody can prove what a scam they are.

Options contracts are exactly this.

The success of such would require extreme precision in modeling assumptions. For example, risk ratings on CDOs with mortgages assumed independence -- but that doesn't hold in reality. At least early[0], on this was recognized as a contributing factor to the misalignment in risk rating, and could have been completely automated.

[0] https://www.wired.com/2009/02/wp-quant/

It takes human cognition to understand the degree of independence. A machine could not have figured out the broader macroeconomic forces at play that led to mass defaults.

Precisely.

Frankly in the financial world there seems to be a market for everything, useful or not.

"In the City, they sell and buy. And nobody ever asks them why. But since it contents them to buy and sell, God forgive them, they might as well."

Are you sure? I've never heard of credit score before moving to the US. Or maybe it's more invisible where I'm from

Reply on news.ycombinator.com