Illiquidity Meets Intelligence: AI-Driven Price Discovery
in Corporate Bonds†
Stacey Jacobsen‡ Kumar Venkataraman§ David X. Xu¶
January, 2026
Abstract
We study the contribution of AI-generated reference prices to intraday price
discovery in markets with infrequent trading. Using corporate bond transactions
and MarketAxess CP+ quotes, we find that CP+ is more informative about
future trade prices than the last trade. Regression results show that CP+ quote
updates reflect market-wide movements in bond, equity, and options markets,
and bond-specific non-public information from the RFQ process. CP+ provides
broad coverage across bonds. Its contributions exhibit a bell-shaped relationship
with liquidity. Following a trade report, CP+ updates quickly in the direction
of the trade, limiting its contribution during transitory price shocks.
Keywords: Artificial Intelligence, Corporate Bonds, Reference Prices, Price Discovery
†We thank Julien Alexandre, Chisom Amalunwez, Hank Bessembinder, Jean Helwege, Burton Hollifield,
Edith Hotchkiss, Xiaowen Hu, Simon Jurkatis, Shuo Liu, Paul Schultz, Susan Thomas, Sinem Uysal,
Jinming Xue, Alex Zhou, and seminar participants at MarketAxess, The Microstructure Exchange, the 14th
Fixed Income and Financial Institutions Conference, NBER Big Data, Artificial Intelligence, and Financial
Economics Conference, SFS Cavalcade Asia-Pacific Conference, and the 16th Emerging Markets Conference
for their helpful comments. We are also grateful to Julien Alexandre, Rick McVey and Sinem Uysal for
sharing MarketAxess CP+ and RFQ data, and to Alie Diagne, Ola Persson, and Jonathan Sokobin for
providing access to FINRA TRACE data.
‡Southern Methodist University. Email: staceyj@mail.cox.smu.edu
§Southern Methodist University. Email: kumar@mail.cox.smu.edu
¶Southern Methodist University. Email: davidxu@smu.edu
in Corporate Bonds†
Stacey Jacobsen‡ Kumar Venkataraman§ David X. Xu¶
January, 2026
Abstract
We study the contribution of AI-generated reference prices to intraday price
discovery in markets with infrequent trading. Using corporate bond transactions
and MarketAxess CP+ quotes, we find that CP+ is more informative about
future trade prices than the last trade. Regression results show that CP+ quote
updates reflect market-wide movements in bond, equity, and options markets,
and bond-specific non-public information from the RFQ process. CP+ provides
broad coverage across bonds. Its contributions exhibit a bell-shaped relationship
with liquidity. Following a trade report, CP+ updates quickly in the direction
of the trade, limiting its contribution during transitory price shocks.
Keywords: Artificial Intelligence, Corporate Bonds, Reference Prices, Price Discovery
†We thank Julien Alexandre, Chisom Amalunwez, Hank Bessembinder, Jean Helwege, Burton Hollifield,
Edith Hotchkiss, Xiaowen Hu, Simon Jurkatis, Shuo Liu, Paul Schultz, Susan Thomas, Sinem Uysal,
Jinming Xue, Alex Zhou, and seminar participants at MarketAxess, The Microstructure Exchange, the 14th
Fixed Income and Financial Institutions Conference, NBER Big Data, Artificial Intelligence, and Financial
Economics Conference, SFS Cavalcade Asia-Pacific Conference, and the 16th Emerging Markets Conference
for their helpful comments. We are also grateful to Julien Alexandre, Rick McVey and Sinem Uysal for
sharing MarketAxess CP+ and RFQ data, and to Alie Diagne, Ola Persson, and Jonathan Sokobin for
providing access to FINRA TRACE data.
‡Southern Methodist University. Email: staceyj@mail.cox.smu.edu
§Southern Methodist University. Email: kumar@mail.cox.smu.edu
¶Southern Methodist University. Email: davidxu@smu.edu
A central function of financial markets is to facilitate price discovery, enabling buyers and
sellers to assess the fair value of assets. This allows them to trade when both sides perceive
a benefit, leading to more efficient markets. Academic research has identified transparency,
defined as observable prices and quantities of completed trades (“post-trade”) as well as
the best available bid and ask quotes (“pre-trade”), as a key market design feature that
supports this process. A recent development in this context is the emergence of Artificial
Intelligence (AI)-based reference prices, generated by algorithmic tools that apply data science
and machine learning (ML) techniques. This article examines the broader implications of
AI-driven models for price discovery in over-the-counter (OTC) fixed income markets.
We focus on the U.S. corporate bond market for several reasons. Unlike equities, which
primarily trade on electronic venues with firm, executable quotes, corporate bond liquidity is
typically provided by dealer firms offering indicative rather than binding quotes. These quotes
are shared selectively with market participants. The market lacks a centralized quotation
system to aggregate dispersed quotes and identify the best available prices. To improve
transparency, regulators have instead focused on improving post-trade disclosure, requiring
dealers to report completed secondary market trades through FINRA’s TRACE system.
Empirical studies have shown that these trade disclosures have reduced customer trading
costs and increased market activity in fixed income markets.1
However, when a long time has elapsed since the last transaction, the information content
of that trade becomes stale. Thus, the effectiveness of post-trade transparency diminishes
when trading activity is sparse.2 In our sample from 2017 to 2023, the average corporate
bond does not report a non-retail trade on 60% of bond-days (see Figure 1, Panel A) and
1See Bessembinder, Maxwell, and Venkataraman (2006), Edwards, Harris, and Piwowar (2007), Goldstein,
Hotchkiss, and Sirri (2007), Schultz (2012), Gao, Schultz, and Song (2017), O’Hara, Wang, and Zhou (2018),
Schultz and Song (2019) and Chalmers, Liu, and Wang (2021).
2Among bonds that do trade, the market is segmented between institutional round lots ($1 million or more)
and retail odd-lots ($150,000 or less). Institutional trades primarily drive price formation, while retail trades,
despite accounting for about 70% of reported trades, are rarely used in pricing benchmarks. Bessembinder,
Kahle, Maxwell, and Xu (2008) recommend excluding non-institutional trades when calculating corporate
bond abnormal returns in order to increase the statistical power of the tests.
1
sellers to assess the fair value of assets. This allows them to trade when both sides perceive
a benefit, leading to more efficient markets. Academic research has identified transparency,
defined as observable prices and quantities of completed trades (“post-trade”) as well as
the best available bid and ask quotes (“pre-trade”), as a key market design feature that
supports this process. A recent development in this context is the emergence of Artificial
Intelligence (AI)-based reference prices, generated by algorithmic tools that apply data science
and machine learning (ML) techniques. This article examines the broader implications of
AI-driven models for price discovery in over-the-counter (OTC) fixed income markets.
We focus on the U.S. corporate bond market for several reasons. Unlike equities, which
primarily trade on electronic venues with firm, executable quotes, corporate bond liquidity is
typically provided by dealer firms offering indicative rather than binding quotes. These quotes
are shared selectively with market participants. The market lacks a centralized quotation
system to aggregate dispersed quotes and identify the best available prices. To improve
transparency, regulators have instead focused on improving post-trade disclosure, requiring
dealers to report completed secondary market trades through FINRA’s TRACE system.
Empirical studies have shown that these trade disclosures have reduced customer trading
costs and increased market activity in fixed income markets.1
However, when a long time has elapsed since the last transaction, the information content
of that trade becomes stale. Thus, the effectiveness of post-trade transparency diminishes
when trading activity is sparse.2 In our sample from 2017 to 2023, the average corporate
bond does not report a non-retail trade on 60% of bond-days (see Figure 1, Panel A) and
1See Bessembinder, Maxwell, and Venkataraman (2006), Edwards, Harris, and Piwowar (2007), Goldstein,
Hotchkiss, and Sirri (2007), Schultz (2012), Gao, Schultz, and Song (2017), O’Hara, Wang, and Zhou (2018),
Schultz and Song (2019) and Chalmers, Liu, and Wang (2021).
2Among bonds that do trade, the market is segmented between institutional round lots ($1 million or more)
and retail odd-lots ($150,000 or less). Institutional trades primarily drive price formation, while retail trades,
despite accounting for about 70% of reported trades, are rarely used in pricing benchmarks. Bessembinder,
Kahle, Maxwell, and Xu (2008) recommend excluding non-institutional trades when calculating corporate
bond abnormal returns in order to increase the statistical power of the tests.
1
