AI And The Municipal Bond Market: Solving The Pricing Problem

For a brief time, I worked for a pricing service, then Thomson Reuters, now LSEG. I took the job for two reasons. First, a colleague I liked worked there, so it would be fun. Second, after joining the chorus of grumblers about how bad municipal bond pricing was, I was curious for a look behind the curtains.

It Pucks

Ask nearly anyone in the muni market about pricing and they usually respond with words that rhyme with “it pucks”. One rationale is the market has what is genteelly referred to as a ‘structural problem’. In broad terms, the problem is that of the $4 trillion par value outstanding bonds with over 1 million unique identifiers, just 0.33 percent of those traded. That was in 2023—a record trading year. Of bonds that do trade, the market is essentially bifurcated between institutional block sizes ($1 million or larger) or retail odd-lots ($100,000 or smaller).

Institutional block trades drive valuations, with trades generally in big new issue underwritings and on the ‘long end’ of the curve: 20 year and 30 year maturities. In most cases, bonds from new issues trade for a few months, then get laid to rest in a mutual fund or SMA portfolio. It’s what market participants call “going to bond heaven”.

Prices on retail odd-lot trades, usually in bonds between 1 and 10-year maturities, are rarely if ever considered in valuations. In part because of that, it’s a woefully inefficient segment of the market. This is peculiar in and of itself, given the ever-increasing record number of trades. Recall that retail odd-lot size of $100,000 and under? In 2023, the average daily customer ‘bought and sold’ par trades in that range was $835.8 million—a 55 percent increase from 2019’s $538.5 million. It looks like 2024 could beat that. But the impact on pricing? Negligible.

[Quick aside. All of the trade and issuance numbers referred to this article can be found in the Municipal Securities Rulemaking Board Factbooks and in the “Market Statistics” links on the MSRB EMMA website.]

On The Job

The first week at the job confirmed my fears. The pricing methodology was rudimentary, at best. Basically, bonds were priced three ways.

First, direct live trade. If there was a trade, that set the price for that bond and all other of those bonds along the 1 year-to-30-year yield curve.

Second was the “comp trade”. While a bond didn’t have a live trade, other bonds that were comparable—issue size, coupon, structure, maturity, rating, and so forth—had live trades. Those “comp trade” prices set the price on bonds that didn’t trade.

The third, the extrapolated trade, is from where the market’s opaque pricing reputation germinates. Since the vast majority of bond issues are under $100 million—in 2023, it was over 90% of the market, a number that has not varied much over the last decade—most of those are structured with individual maturities under $1 million. Correspondingly, that same vast majority of bonds don’t have and never will have institutional block size trades to get pricing guidance from.

However, any bond held by a regulated investment advisor for a client—mutual fund, SMA, broker/dealer—needs marked-to-market, end-of-day pricing to comply with the “Fair Value” provisions of SEC Rule 2a-5. But as noted, over 99% of these bonds lack live trades. To solve for this, pricing services create complex matrices to determine those values. Billions of dollars of bonds are priced off of a quantitatively derived educated guess off of what may be stagnant data of varying or indeterminate accuracy.

It’s a methodology, perhaps done in good faith, that generates a value number. But is that value truly a fair market value if the price you see on your brokerage statement isn’t likely to be even close to the price you’ll get if you want to buy or sell that bond?

This is not to suggest that data on bonds that do trade is not readily available and current. The Municipal Securities Rulemaking Board offers a real time trade data subscription. Additionally, the MSRB’s EMMA (“Electronic Municipal Market Access”) offers a price discovery tool, a regularly updated most actively traded bond list, and other price resources on the website. It’s quite impressive.

But for the vast number of bonds just quietly accruing interest, their pricing doesn’t come until the end of the day. There is no up to the second, or even up to the minute pricing on bonds. All this only reinforces the market’s reputation for opaqueness.

Look closely enough at the opacity and one thing does become a bit clearer. Municipal bond pricing, as it is now, is more akin to a big echo chamber: a few trades reverberate throughout the market.

As my days at Thomson wore on, I drifted over to the more intellectually challenging valuations on high yield, distressed, and defaulted bonds. It was a brief stay; I don’t think I lasted a full year before leaving to join the muni fintech startup Neighborly.

Yield to the Curve

As if pricing on individual bonds wasn’t opaque enough, it extends to the yield curve.

Every fixed income market professional tracks the yield curve. The yield curve is comprised of the rates on bonds for each maturity from 1 year to 30 years. In the municipal bond market, the U.S. Treasury yield curve and the tax-exempt AAA yield curve are the two yield curves followed most closely.

As fixed income investments, municipal bond prices are based off of these curves—the “spread”, as it is referred to. Spread is the yield on the bond minus the yield on the municipal triple-A yield curve. For example, if a 10 year maturity bond yield is 3.75% and the same maturity on the triple-A curve is 2.75%, the spread is 100 basis points (each basis point being 1/100^th of 1%).

Which curve is used (and it may be one or both) depends on the circumstances, but either way they are a traditional core component of bond valuation. If interest rates go up or down, it is reflected in the curve. Bond values priced off of the curve are adjusted accordingly. For the municipal curve, there are other considerations as well, such as state tax rates or credit ratings. But for a rule of thumb, if you’re buying or selling a bond, one way to determine a generic price would be to see where the rate on the yield curve is for that bond’s maturity and then pricing the bond off of that.

But here’s the rub: what if you don’t have an accurate, up-to-the-minute yield curve to refer to?

Traditionally, the market has turned to Refinitiv’s Municipal Market Monitor (TM3) yield curves if only because there haven’t been many other competitors. Derived from longer maturity AAA state general obligation bonds, the TM3 Municipal AAA MMD has been the market’s long-time but unofficial benchmark. It reflects where institutional buy side (i.e., mutual funds) and sell side (broker/dealers) expect to transact. It is released when something material changes in the market, which is usually three times a day, but the timing can be variable. TM3 has a library of curves for subscribers, boasting some 250 survey-based scales (read: yield curves) across numerous sectors, credit rankings, and other market matrices, many with decades of yield curve data.

But between these postings, investors, underwriters, traders, and anyone else involved in trying to price a bond accurately, are left without an up-to-the-minute benchmark yield curve. Curves other than MMD do exist (you can find those on the MSRB website) but most are subscription-based and are publicly available only at the end of the day. Not particularly helpful in a fast-moving market situation.

The market needs a solution.

The Tech Calvery Arrives

Having worked with data scientists and software engineers, I can tell you firsthand that what turns them on the most is a good challenge. The more intractable the problem, the more impossible it seems, the more they are drawn to it. It’s like magnets to iron.

Moreover, they see the old guard of TM3, ICE, and Bloomberg as exposed and vulnerable to newer, faster, and more efficient AI technologies. Combine a multi-trillion dollar pricing problem with the opportunity to upend embedded incumbents? It’s a nearly irresistible tech challenge.

Meet the Startups!

Enter four very eager and AI-driven fintech start-ups fast to use data science, artificial intelligence, and machine learning to tackle the muni market’s pricing and yield curve problem. And that’s four just as of today. It’s likely there will be more.

Startup #1: Spline Data

Spline Data creates real-time, model-driven pricing and yield curves for the municipal bond market. Utilizing around 140 metrics from trading and other data, the firm develops cutting-edge statistical analyses to create real-time yield curves and predictive pricing models by applying AI methodologies. Measuring and benchmarking its pricing and curve performance against actual market movements and trade prices, the firm focuses on predicting execution prices—“nowcasting”, not “forecasting”. Backtesting for variances and updating its models accordingly, the firm’s machine learning is always growing and building for greater accuracy.

[Fun Fact: “spline”—with the “i” pronounced as “eye”—is defined as “a piecewise polynomial function used to approximate a smooth curve in a line or surface” or, for those of us who are mere mortals, it is the math applied to a data set to create a continuous and irregular curve, like a Yield Curve.]

Spline’s founder, Matthew Smith, started Spline Data in 2022 after serving as the Head of Trading at Headlands Tech Global Markets, later acquired by TD Securities. The implications of more accurate pricing in a timelier manner will have a waterfall effect throughout the market, contends Smith. Yes, automation is the far-and-away leading benefit of tech and AI integration, but the effects quickly cascade—better relative-value identification, a reduction in lead-time for building algorithms, transaction cost analysis, best execution, tax loss harvesting, portfolio optimization without ever initiating an RFQ. The list goes on…and on.

His dream outcome from all this are tighter bid-ask spreads, better execution in odd-lot trades under $1 million and, in the end, an overall more liquid and efficient market benefiting issuers and investors alike.

He is not alone.

Startup #2: ficc.ai

The principals of ficc.ai (the company’s name is in lowercase) realized that given the size and illiquid nature of the market, no human or group of humans can offer real-time pricing. Hence, the firm is solely focused on providing customers accurate and real-time price evaluations. The solution—AI models, learning from tens of millions of data points—offers accurate real-time pricing for the entire universe of municipal bonds, enabled by the latest advances in machine learning. Applying neural network architecture, these ensemble models are able to learn interactions between subsets of features to provide an accurate price. Along with terms and conditions (TNC) data as well as current trading data for every bond, the firm uses prices of Muni ETFs, changing up to the second and throughout the day, to capture the tone of the market at any given point in time. The firm ingests roughly 1TB of data per day. Consistent backtesting and rigorous analysis to evaluate the accuracy on hundreds of features ensure price accuracy across the entire universe.

Startup #3: SOLVE

SOLVE isn’t exactly a startup (it was founded in 2011), but it sure acts as innovative as one. Through both organic growth and acquisition, the firm positioned itself as a leading market data platform provider for fixed-income securities. Already robust with datasets of real-time bids, offers, covers, and other market information on a variety of fixed income instruments, from corporate bonds to syndicated bank loans, its acquisition of Lumesis in 2022 expanded its reach into the municipal bond market. The beta launch of SOLVE Px™ offers AI-generated predictive trade levels on over 900,000 municipal bonds. As SOLVE Co-Founder Eugene Grinberg put it, the platform seeks to see what the next trade level will be with a goal of minimizing prediction error. From his perspective, besides more realistic and real-time market price information, AI also offers the benefit of objectivity, providing an unbiased projection of the next trade price.

In addition to bond reference data—coupon, maturity, and so forth—the firm extracts hundreds of millions of quotes from both structured trade data and unstructured quote data parsed from electronic messaging between traders. Additionally, with their obligor database (economic and demographic data mapped to the CUSIP level), they will consider utilizing such nonfinancial data points where they provide meaningful insight. Applying proprietary AI methodologies to 300 features, SOLVE parses out valuation drivers and, in as close to real-time as possible, generates bond prices.

Grinberg hopes to increase transparency and reduce unnecessary variability in pricing. Like others, he believes the market doesn’t operate in a fair and transparent fashion. Applying data and AI, Solve seeks is to level the playing field between institutional and retail investors.

Startup #4: 7 Chord

7 Chord is the newest entrant to the municipal bond pricing universe. Based on her 20-years of credit trading and market structure experience at Blackrock and other Wall Street behemoths, Kristina Fan saw first-hand that rules-based approaches in bond valuation were too crude and rigid. They often did not produce accurate results, especially during volatile markets. Seeing the pricing problem as a way of turning sci-fi ideas into industry-changing products (her words, not mine) she founded 7 Chord in 2016 with Roy Lowrance, a veteran technologist who at the time ran the Center for Data Science at NYU. With markets evolving with ever increasing speed, 7 Chord believes it is more important than ever that prices of investable assets reflect all the information available up to this moment in time. The firm sees AI’s ability to process vast amounts of data in real-time to create dynamic price adjustments based on current market conditions. The municipal bond market is now targeted in their sights.

Its solution, BondDroid, is essentially an orchestra of over 400 different pricing bots. Each has a unique methodology and, using hundreds of different features, is designed to mimic a certain market persona or regime. Some bots are simple, some very complex, but overall it’s these diversification of approaches that feeds an AI engine continuously retraining itself in real-time to produce a ​market-relevant price estimate. That diversification is particularly necessary in the heterogeneous fixed income markets. Moreover, AI implementation is more flexible than any rules-based method and is faster and more scalable than any human.

In her view, probably the most important positive effect of AI pricing is the emergence of this systematic accuracy measurement methodology, as well as a more precise definition of what the ground truth is with regards to pricing of thinly traded assets. Currently, it’s far from perfect. More work is needed for the industry and the regulators to perfect and adopt these methods more broadly.

The royal we agree to all of the above.

The Future was Yesterday

With all this predictive analytics, real-time pricing, yield curves, and spreads, the prediction soon becomes the reality. If you’re sitting on your trading desk and you have a solid idea at what level the next trade is coming, you’re not bidding or asking at the existing trade, you’re bidding or asking at the anticipated trade price. In fact, you’re not human. You’re likely a computer algorithm that is being trained by a neural network to trade against other computers also being trained by a neural network.

Sounds vaguely like a movie I saw back in the ‘80s.

The next article in the AI and the Municipal Bond Market series will cover fundamental credit analysis. With AI, is it even relevant? Or can AI prove its relevance?