Will AI put bond fund managers out of work?

aroominyork

Managers of active bond funds ask two basic questions: what is the risk of this company defaulting, and what does the yield tell us the market thinks is the risk of default? If the manager thinks the market overstates the risk, he is a buyer. This is essentially a mathematical exercise; I doubt fund managers sit down with companies, as equity fund managers do, to be sold the story of future success.

This seems a process where AI could do a lot of the work and lead to the yield on bonds accurately reflecting their risk, driving active funds out of business.

tacpot12

I think it could drive out some active managers, for the reason you sggest, but humans are still required; to set the aims of the fund.

I'm yet to see any evidence that AI can imaging what will sell in any particular marketplace, e.g. can AI learn how to figure out what a customer might want. Can AI figure out what the problems with the current products are and how to improve them? Given that there are a near infinite number of choices you could make to improve a bond fund, e.g. adding chocolate chips, can AI figure out what might matter to human customers? How could it do that if there is no data for it to be trained on because the product feature or concept hasn't been invented yet?

aroominyork

You wouldn't need to work out what would sell. You just offer index funds with combinations of credit quality, duration and geography (and maybe hedged/unhedged), all AI-priced to correctly reflect their risk, and there will be enough takers to make all those funds viable. There are many fewer permutations than with equities.

masonic

A mathematical exercise that involves building a model that ranks securities according to multiple factors sounds like something we've had available for about 15 years. There still seems to be plenty of active managers around using their intuition and judgement, then being paid handsomely for it.

In fact it seems like these approaches have gained more traction in equity funds, where active management seems to have a poorer record in general.

aroominyork

It's the binary nature of bonds - will or won't they default - that makes them seem a natural for a mathematical approach. Is there any analysis of these models' accuracy of outcome over the last 15 years?

masonic

https://forums.moneysavingexpert.com/discussion/comment/82157391#Comment_82157391

I think the challenge is that default rates are so low, meaning that factors other than avoiding defaults are required to beat your peer group. For investment grade bonds, annual default rates have averaged less than 0.2% per year since 2011. A simple model that always predicted a bond would survive the next year would be 99.8% accurate, which is tough to meaningfully improve upon. This is the Accuracy Paradox.

High yield has a recent average default rate of 3-4% per year. Losses from defaults would not be total, perhaps you would recover roughly half. Meaning an equal weight junk bond total index fund would suffer an average of 1.5-2% drag on returns per year as a result of defaults. Currently an active manager can perhaps avoid a decent chunk of that while maybe charging a 0.5% premium on the AMC. So a really good model could buy back most of that 0.5% in efficiency savings if avoiding defaults was all an active fund manager thought about when selecting bonds.

I think we both know that active managers exist who have at various times delivered much higher returns than could be explained by just avoiding defaults over the period under examination. It is human nature to chase such performance and such managers. While an "AI manager" could be marketed to investors (it would need more behind it than it simply learns to avoid defaulting bonds), and it might appeal to some investors, it likely won't appeal to all, and will probably make mistakes just like human managers make mistakes.

aroominyork

You're missing my point. I am not saying a good model could consistently match or outperform the best managers. I am saying a model which consistently predicted the likelihood of default in the binary world of bonds could lead to something close to perfect pricing. There would no longer be an edge for Jonathan Golan to capture.

masonic

https://forums.moneysavingexpert.com/discussion/comment/82157417#Comment_82157417

Why would it lead to something close to perfect pricing? The model of assuming no bonds default is already a very good model. Models are generally not good at extrapolation, so I doubt anyone could have total confidence in a model with even perfect recall of historical data.

If there is news that a company is in deep trouble, but an AI predicts it will survive, do you think institutions with millions invested in the corporate debt will relax and hold their positions? Or flood into it if its credit was already priced adversely?

aroominyork

If there is news that a company is in deep trouble, but an AI predicts it will survive, do you think institutions with millions invested in the corporate debt will relax and hold their positions?

Say 20 companies in the previous year had been in similar trouble and AI had correctly predicted, with reasoning, they would all survive. Would institutions begin responding differently? I expect so. Anyway, let's hope it doesn't come to this and my bond funds have the opportunity to continue outperforming.

masonic

https://forums.moneysavingexpert.com/discussion/comment/82157474#Comment_82157474

In that case it wouldn't have beaten my model ;)

Now lets say 10 of those 20 companies did go to the wall and on average bondholders lost 50%. The model correctly predicted all 10. Then that would be more remarkable, as that could represent (using the junk bond default stats) an edge of about 2% for those who would otherwise have done nothing. However, those who disinvested would have lost less, perhaps they'd have suffered adverse pricing of 10-20%, resulting in about a 1% drag on returns on their diversified portfolio. Then there are those who avoided being invested in the questionable bonds before the trouble broke out.

But I don't think, even given all publicly available corporate information going back decades, you can build an AI model that would predict a company's future liquidity with sufficient accuracy to significantly influence the market. Maybe a model released today that is highly accurate over the next decade would start to gain traction over the following decade. But a track record is no guarantee.

It could be rather perverse that, such a model may be best suited to determine whether AI circular financing is going to blow up, since it could be given highly relevant inside information. But I don't think the AI hyperscalars would want that.

Vitor

AI will almost certainly make credit analysis cheaper and faster, and may reduce the scope for managers whose edge is mainly processing publicly available information. But if everyone uses similar models, any advantage gets competed away rather than producing "perfect pricing". The remaining opportunities will tend to be where the inputs are uncertain, markets become illiquid, management behaviour matters, or the model is wrong.

Every model looks wonderfully precise until the assumptions stop matching reality. Black-Scholes is the obvious cautionary tale: enormously useful, mathematically elegant, but dependent on assumptions about volatility, liquidity and market behaviour that can fail badly in stressed conditions.

So I could see AI squeezing active bond management fees and headcount quite hard. Putting active managers out of business altogether seems a much bigger leap.