For an industry historically averse to change, the municipal market is slowly beginning to incorporate artificial intelligence into workflow systems, tools that succinctly summarize documents, such as official statements, as well as internal and external communication systems.
But it is far away from broad acceptance for an industry dependent on human interaction to get business done.
In the next few years, firms expect to use it to respond to written requests for true interest costs on deals, write first drafts of documents, create computer programs and analyze credits, and more.
Professionals in the industry anticipate widespread but gradual adoption in coming years.
Several municipal
What is now TD Securities Automated Trading emerged from hedge fund Citadel about eight years ago. Citadel used machine learning to analyze equities, said Matt Schrager, co-head of TD Securities Automated Trading.
Schrager said his group has used machine learning, first as Headlands Tech Global Markets and now as TD Securities Automated Trading, to determine fair bid and offer prices for munis and other fixed income instruments.
As a largely buy-and-hold asset class, most of the $4.1 trillion outstanding universe of individual municipal bonds trade very infrequently. Given this, price discovery is challenging in the muni market, Schrager said.
Since 2015, Schrager said, his group has used
To determine fair value for a bond that hasn't traded in weeks or months, "we need to correlate that bond to similar bonds that have traded recently," Schrager said. "However, no two bonds in the market are quite alike, so we must be careful in this analysis to account for relevant differences."
Machine learning "helps us perform this kind of analysis in a rigorous way that both incorporates as much information as possible while also accounting for important differences between securities," he added.
Earlier this month, market data platform SOLVE
AI offers value in muni pricing, Lisa Schirf, managing director at Tradeweb, said.
"Recent advancements using AI, including Tradeweb's Municipal Ai-Price, which publishes intraday and end-of-day closing prices, have made accessing reliable and quality pricing levels easier and more efficient for municipal bond market participants," she said.
Spline Data, a quantitative municipal market data firm founded in May 2022, uses a substantial amount of machine learning, a subset of AI, in its operations, said firm CEO Matthew Smith.
"We use machine learning to accomplish anything from classifying groups of similar bonds to producing yield estimates for our curve based on trading data," Smith said.
"Spline uses a fairly popular large language model (LLM) [artificial intelligence program] to assist us in writing code for our product," Smith said. "Oftentimes, what are relatively mundane and simple coding tasks can be outsourced to an LLM. For example, if we are building a plotting function to visualize our data, we may prompt the LLM to produce a chart of what we want to compare and that's done in 30 seconds as compared to 30 minutes" if Smith wrote the program for a comparison.
A large language model, such as Open AI's GPT-4, is an algorithm that can summarize huge data sets and create content.
Munichain CEO Matthew Gerstenfeld said his company plans to use artificial intelligence for some customer interactions.
In its
Customers will be able to speak to a human, if they choose, Gerstenfeld said, since that will be more appropriate for some inquiries.
Despite working to incorporate AI, Gerstenfeld said Munichain is still exploring its use.
Ken Hoffman, president of DPC Data, and Schrager said, Open AI's introduction of ChatGPT in November and subsequent upgrade with GPT-4 in March will impact the industry.
"Generative AI, such as Open AI's ChatGPT, has shown that new content can be created based on data itself and not any predefined rules," Hoffman said. "While it learns and can be trained by humans to create content based on examples and patterns, the results are more creative."
While he wouldn't commit to how AI will impact the muni market, Hoffman said, "it is conceivable that it will create content that is truly new and original," including "investment suggestions, credit analysis and scoring, pricing evaluation, etc."
DPC Data is investigating potential uses "for Generative AI, with an eye toward leveraging our extensive credit reference and financial database into new, innovative solutions for the municipal market," said Vice President Triet Nguyen.
Spline's Smith expects some integrated development environments (IDE) tools to be available "within the year but for custom implementations, I'd guess closer to two to three years." Artificial intelligence IDE tools will make computer programmers and quantitative analysts much more efficient when developing computer code, he said.
Hoffman said others in the industry claiming to use AI are "using advance algorithms to output results that humans then verify or correct. It is helpful technology, but not a technology that is able to generalize data on its own to produce meaningful, creative content."
Gerstenfeld said Munichain started work on Munichain web application in 2022 and it remains in a testing stage. The firm expects to further integrate AI into its operations gradually, perhaps over the course of 10 years. For the time being, employees and computers serve customers.
Schrager and others said many firms are approaching ChatGPT cautiously due to concerns about data privacy and protection. Schrager hopes each firm will eventually have its own integration with ChatGPT and data won't be shared between firms.
ChatGPT allows non-programmers to do things that, in the past, "would have required a software developer," he said. "This will increase the productivity of non-technical market participants, like traders, portfolio managers, credit analysts, etc."
As an example, Schrager showed how the program summarized a more than 100-page bond official statement based on a written request. "Prior to ChatGPT, producing that table … would have required either A) manually reading through a 100-plus page official statement, or B) writing custom code to do it for you."
Other examples of what ChatGPT could do for muni participants, Schrager said, include compiling and displaying information about population size and tax revenue for issuers, summarizing credit research, and cross-referencing a fund's muni inventory with offers available to produce a list of offers where the price is cheaper than the ICE evaluation. ChatGPT should be able to complete these tasks in the future.
The sort of in-house artificial intelligence that TD Securities has will affect the muni industry more than products like ChatGPT, Schrager said.
Hoffman said ChatGPT could look at data and generate investment suggestions, credit analyses and scoring, and pricing evaluation. "Will that content be better or worse than the content generated by the human mind? Due to the amount of data that the AI model can process and learn from, it is likely" to be better.
Data is going to become "more abundant and accessible to market participants [and] so the data realm is likely the first area of munis to be impacted by AI," Smith said. "Machine learning in practice has proven to be a substantial source of edge for market participants and effectively acts as a performance multiplier for existing participants. Despite this, it takes a substantial investment of time and capital to develop, maintain, and improve machine learning models internally so companies will continue to benefit from utilizing outside data providers.
"As for referential data, optical character recognition coupled with a fine-tuned large language model could very likely lead to data sets in this space becoming more commoditized," Smith said.
"For the sales, trading, and research side of the market, firms are already using more primitive natural language processing capabilities to capture both qualitative and quantitative information from communications and disclosures, so it's not a huge leap to assume that will continue and potentially even accelerate with the introduction of large language models," Smith said.
Bond counsel Orrick was an early adopter of technology, said Justin Cooper, co-head of public finance at the firm. He said he anticipates using artificial intelligence to summarize and write first drafts of documents.
Some said muni issuers may use artificial intelligence in the future. Cooper said many issuers' chief financial officers are being asked to do more with less personnel than in the past. Artificial intelligence can be beneficial in these situations, he said.
A constant drumbeat from market participants is they have concerns that AI may lead to layoffs: technology replacing humans.
Cooper said he does not expect that in the near future, since clients hire Orrick for its institutional market knowledge and its lawyers' judgment.
While that sentiment is echoed by Smith and Hoffman, Smith said the number of evaluators who determine secondary prices may shrink.
Amid the growing excitement and hype about artificial intelligence, DPC's Nguyen offered words of caution: "It should be noted that the term 'AI' has historically been misused in our market to refer to semi-automated processes with significant human intervention. Without a self-learning component, that's really not 'AI'."
As for the future, Schrager said, the many muni firms that haven't embraced artificial intelligence will start to do so. "The train will continue to move in the direction it is going."
