As the muni market becomes more comfortable integrating artificial intelligence and machine learning in areas including workflow systems, data and credit analysis and compliance, AI still struggles in various muni functions. One such area is pricing high-yield bonds, since fewer data points are available for AI models to generate prices, market participants said.
In general, high-yield bonds are difficult to price using AI, whether "it's a regular algo at an evaluated pricing shop or it's an AI-based approach from some of the newer AI-related models that are out there," said Michael Pellerito, SVP of innovation at SOLVE.
For an AI model, when "you're trying to price high-yield, you're training your model based on other high-yield bonds and other high-yield traits," he said.
And with high-yield being only 7% of the market, the remaining 93% of the market, which is investment-grade, might be "thrown away" because "you're trying to ... get training information for these high yield bonds, there's just so much less of it," Pellerito said.
Due to difficulties, J.P. Morgan subsidiary PricingDirect does not price high-yield bonds in its AI evaluating pricing model. The firm prices only investment-grade munis using an AI and machine learning model, said Daniela Durao, executive director of PricingDirect, during an AI-driven municipal bond pricing webinar on Tuesday.
However, high-yield munis are too sparse and illiquid and there's too much spread differentials to currently run an AI model, she said.
Fewer data points to train an AI could result in a "degradation of accuracy," and with high-yield representing a small part of the muni market, there may be a little bit more of an error uptick for high-yield, Andrew Hui, senior product manager at MarketAxess, said at the webinar.
Compared to investment-grade, which is pretty stable, high-yield is more credit-centric. Due to that, the credit characteristics of any bond can change at any moment, and that makes it a little bit more volatile, Pellerito said.
Another problem with high-yield is that trading is "very bursty," with a lot of information at once and then nothing, said Matt Smith, founder and CEO of Spline Data.
There's just not a lot of outstanding high-yield debt, so it's not as easy to "impute " what other high-yield sectors are doing, he said.
In high-yield, dirt bonds might move independent of transportation bonds and they're not spread products to some degree, Smith said.
"That's where the most issues come from. It's just a lack of information, a lack of comparability," he noted.
Unlike PricingDirect, both Spline Data and SOLVE use AI and machine learning to price high-yield bonds.
For the former, Spline Data uses AI to produce three curves — a AAA curve, an individual bond predictive pricing model and a primary market model — while SOLVE uses AI and machine learning for predictive pricing.
Spline Data has a base model that acts like a trading desk. There are three different models, and a so-called head of a desk that allocates weights. All of that initially was geared toward investment-grade, because Spline's customers are algos trading IG, Smith said.
However, as the firm has added customers, especially on the buy side, they'll say, "Hey, your pricing seems off on this type of bond." And if Spline gets enough of those inquiries, the shop can start to define pockets of bonds that it's not very good at pricing and create a model specifically for that part of the market, he said.
"If [the model] is any good, then that optimizer gives all the weight to that model and that pocket of the muni market," Smith said.
As for SOLVE, the shop will continue to improve the accuracy of its pricing for high-yield by adding signals and bringing in more credit-centric-related signals.
To improve AI-driven high-yield pricing, more data and trades would help, Smith added.
"As the models get better and better, you're going to see better and better pricing [for high-yield], and that error is going to be reduced," PricingDirect's Durao said.
