Reading robots
One amusing and conceptually interesting area of financial research involves having a robot read issuers’ public documents—bond offering documents, corporate annual reports, press releases, earnings-call transcripts, etc.—to see if it can extract any market-beating signals from the documents. It is conceptually interesting because it is a particularly stark challenge to the efficient-markets hypothesis: If there is one set of information that every investor has, it is these public disclosures, so if there is information in those filings that is not incorporated into the market price then something has gone rather wrong with market efficiency.

It is amusing because it is very easy to explain what has gone wrong: Nobody reads these documents, because they are boring. Your market-beating edge can be actually sitting down and reading a boring prospectus cover to cover, says science. (Warren Buffett agrees!) It is also amusing because the signals tend to be a little silly; it’s all stuff like, you know, if the chief executive officer makes a good joke on the earnings call then that’s good for the stock or whatever.

Anyway here’s a story about how Morgan Stanley research strategists “experimented with farming out the job of reading 120,000-word bond prospectuses to robots, seeing if the results could yield a sort of CliffsNotes that may separate the signal from the noise.” Specifically they had the robots read official statements for certain kinds of risky municipal bonds, and they found some useful signals, including:

Executive bios: Speculative developments tend to rely on the word “Mr.” to highlight the management of the project, since the riskier deals need to play up their executives’ skills as a key selling point.

Boring is better: Higher-quality debt tended to have more references to the financial statement than defaulted or downgraded debt. The more “boring” the documents, the better, the strategists said.

“The more boring the better” is pretty good advice for corporate lawyers really.

This is interesting but, as an occasional reader of bond documents, I was mostly struck by what the robots aren’t doing. One of my favorite recent stories in bond documents is the one about how private equity firms changed a sentence in bond documents, which usually says that companies can’t make restricted payments if “a Default or Event of Default shall have occurred,” to instead say that they can’t make the payments if “an Event of Default shall have occurred.” Robert Smith explains:

To a layman, this sounds the same. But there is an incredibly important distinction.

If you miss an interest payment, it's a default. But the formal "event of default" doesn't happen until 30 days later, as the bonds have a grace period.

By deleting the default part, the company could miss an interest payment, and then strip a load of cash out of the business before formally defaulting. That is really bad!

Bondholders, as it happened, noticed the change and pushed back in a deal last year—but they hadn’t noticed it in several previous deals. So now there are bonds out there that allow their issuers to stop paying interest and then strip a bunch of cash out of the company to pay their private-equity owners before the bondholders can do anything. And on the one hand after the fact one can scold the bondholders and say “too bad, you have to read the documents carefully,” but honestly you are tired and have a lot of work to do and it is understandable that you’d miss a change from “a Default or Event of Default” to “an Event of Default.” Or you could scold the issuers, and their private-equity owners, for pushing this change and pulling a fast one on the market, but they are just trying to maintain flexibility and it is understandable that they would push for as much optionality as the market will give them. Or you could scold the underwriters, which is what I would do—as a former underwriter myself, I absolutely believe that it is the moral duty of an underwriter to push back against weird squirrelly non-market terms in bonds, or at least to prominently flag them to investors—but this is tricky, and there are those who believe that underwriters are supposed to work for the best possible outcome for issuers rather than for a fair outcome for the market.

Maybe there is a solution to be found in technology? In a sense the ideal reader of a tedious hundred-page bond document is a robot. The robot reads fast, it doesn’t get bored, it doesn’t skim over parts it’s read a thousand times before and miss subtle changes. If there were a robot trained to spot subtle changes in bond documents then that would be super useful.

But that training is hard! It is hard for the robot to know what changes in bond documents are meaningful, because (the robot does not actually possess general intelligence and) there is no real training set, no effective way to correlate “three words were deleted here” with “in three years the private-equity owners might strip cash out of the company and we won't be able to do anything about it.” (I mean, until it happens; then you’ll have the data, but then it’s too late.) The robots can use the documents to analyze the quality of the issuer, but they can’t actually use the documents to analyze the quality of the documents. Reading every word of a boring bond document to figure out if it’s going to hose investors is still, strangely, a job only humans can do.