Financial Analysts Journal

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/ufaj20

Predicting Corporate Bond Illiquidity via Machine

Learning

Axel Cabrol, Wolfgang Drobetz, Tizian Otto & Tatjana Puhan

To cite this article: Axel Cabrol, Wolfgang Drobetz, Tizian Otto & Tatjana Puhan (2024)
Predicting Corporate Bond Illiquidity via Machine Learning, Financial Analysts Journal, 80:3,
103-127, DOI: 10.1080/0015198X.2024.2350952

To link to this article: https://doi.org/10.1080/0015198X.2024.2350952

© 2024 The Author(s). Published with

license by Taylor & Francis Group, LLC.

View supplementary material

Published online: 24 Jun 2024.

Submit your article to this journal

Article views: 497

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ufaj20
Financial Analysts Journal | A Publication of CFA Institute Research
https://doi.org/10.1080/0015198X.2024.2350952

Predicting Corporate Bond

Illiquidity via Machine
Learning
Axel Cabrol, CFA, Wolfgang Drobetz, Tizian Otto, and Tatjana Puhan
Axel Cabrol, CFA, is Co-Deputy CIO at TOBAM, Paris, France. Wolfgang Drobetz is a Professor of Finance at the Faculty of Business
Administration, University of Hamburg, Hamburg, Germany. Tizian Otto is a Postdoctoral Researcher in Finance at the Faculty of Business
Administration, University of Hamburg, Hamburg, Germany. Tatjana Puhan is Head of Asset Allocation at Swiss Re Management Ltd, Zurich,
Switzerland and Adjunct Faculty Member at the Faculty of Business Administration, University of Mannheim, Mannheim, Germany.
Send correspondence to Tatjana Puhan at [email protected].

This paper tests the predictive per- growing strand of literature suggests that machine learning can
formance of machine learning meth-
ods in estimating the illiquidity of
US corporate bonds. Machine learn-
ing techniques outperform the his-
A enhance quantitative investing by uncovering exploitable non-
linear and interactive effects between predictor variables that
tend to go unnoticed with simpler modeling approaches (see Blitz
et al. 2023, for an excellent review of machine learning applications
torical illiquidity-based approach, the in asset management). The majority of these studies use machine
most commonly applied benchmark learning techniques to predict stock returns, applying a large set of
in practice, from both a statistical predictor variables. Most prominently, Gu, Kelly, and Xiu (2020) and
and an economic perspective. Freyberger, Neuhierl, and Weber (2020) show that machine learn-
Gradient-boosted regression trees ing–based approaches outperform linear counterparts and generate
perform particularly well. Historical remarkably high Sharpe ratios (of about 2 or even higher).1 Bianchi,
illiquidity is the most important sin- Bu€chner, and Tamoni (2021) and Bali et al. (2022) confirm the effec-
gle predictor variable, but several tiveness of machine learning techniques in predicting government
fundamental and return- as well as and corporate bond returns, respectively. Nevertheless, compared to
risk-based covariates also possess the literature related to equities, machine learning applications in
predictive power. Capturing nonlin- fixed-income research have received much less attention. This gap in
ear effects and interactions among the literature may be explained by the fact that our understanding of
these predictors further enhances the risk–return tradeoff is still less developed in bond markets than
forecasting performance. For practi- in stock markets (Dickerson, Mueller, and Robotti 2023; Kelly,
tioners, the choice of the appropri- Palhares, and Pruitt 2023). We contribute to this recent literature by
ate machine learning model depends testing the predictive performance of machine learning methods in
on the specific application. estimating the expected illiquidity of US corporate bonds.

Keywords: corporate bonds; bond

The authors thank Daniel Giamouridis (the journal's associate editor), two anony-
illiquidity; quantitative credit research;
mous referees, Yakov Amihud, Maxime Bucher, Tristan Froidure, Patrick Houweling,
illiquidity forecasting; machine learning
Harald Lohre, Robert Korajczyk, Daniel Seiler, Michael Weber, and the participants
of the Mannheim Finance faculty seminar for insightful suggestions and remarks.
This is an Open Access article distributed under the terms of the Creative Commons
Disclosure: No potential conflict of interest Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
was reported by the author(s). unrestricted use, distribution, and reproduction in any medium, provided the original
work is properly cited. The terms on which this article has been published allow the
posting of the Accepted Manuscript in a repository by the author(s) or with their
PL Credits: 2.0 consent.

Volume 80, Number 3 © 2024 The Author(s). Published with license by Taylor & Francis Group, LLC. 103
Financial Analysts Journal | A Publication of CFA Institute

In comparison to actively managed stock portfolios, we use a large universe of US corporate bonds,
there is limited alpha upside in a classical bond port- three illiquidity measures that capture different
folio case, where individual bonds are bought as aspects of illiquidity, a broad set of machine learn-
cheaply as possible and then often held until maturity ing–based illiquidity estimators, and a comprehensive
(hoping that no default occurs). In this setup, every set of predictor variables based on historical illiquid-
basis point in transaction cost savings is crucial for ity, fundamental predictors, return-based predictors,
the success of such a strategy. Considering that cor- risk-based predictors, and macroeconomic indicators
porate bonds are an asset class that is inherently to uncover exploitable nonlinear and interactive pat-
plagued by illiquidity, the scarcity of work on predict- terns in the data.
ing corporate bond illiquidity is surprising. Although
Historical illiquidity is the most commonly used
the question of how to generate outperformance is
benchmark for predicting future bond illiquidity in
the most important one for every investor, any out-
the asset management industry. Examining level fore-
performance potential depends on whether a seem-
casts of illiquidity, our results confirm that machine
ingly superior trading strategy can be efficiently
learning–based prediction models that incorporate
implemented in practice. Therefore, only a reliable
our comprehensive set of predictor variables outper-
estimate of a bond’s future liquidity enables an inves-
form this popular benchmark. Tree-based models and
tor to assess whether this bond is priced in line with
neural networks, which additionally allow for nonli-
its fundamentals and to convert the return signals
nearity and interaction effects, perform particularly
into a profitable investment strategy after accounting
well. For example, compared to the historical illiquid-
for transaction costs and other implementation fric-
ity benchmark, the average mean squared error
tions. Moreover, under the so-called SEC Liquidity
(MSE) is more than 23% lower for neural networks.
Rule, accurate predictions of future bond liquidity are In a statistical sense, based on the Diebold and
essential from a regulator’s and financial market Mariano (1995) test, neural networks outperform the
supervision perspective to monitor bond funds' benchmark in more than 87% of the sample months.
liquidity risk management. In addition, they are in the so-called model confi-
The objective of our paper is to capture the rich fac- dence set (Hansen, Lunde, and Nason 2011), that is,
ets of illiquidity in corporate bond markets using they are among the best-performing models that sig-
machine learning methods. Most studies that feature nificantly dominate all other forecast models, more
elements of bond illiquidity predictions rely on than six times as often. We attribute these improve-
Amihud’s (2002) AR-1 approach (Bao, Pan, and Wang ments in prediction quality to the inclusion of slow-
2011; Friewald, Jankowitsch, and Subrahmanyam moving bond characteristics, such as age, size, and
2012; Dick-Nielsen, Feldhu €tter, and Lando 2012; rating of a bond, as predictors as well as the ability
of both tree-based models and neural networks to
Bongaerts, de Jong, and Driessen 2017). However,
incorporate patters of nonlinearity and interactions in
illiquidity, particularly in a complex market such as
the relationship between expected illiquidity and
the one for corporate bonds, is multifaceted and
these predictor variables. Furthermore, forecast
incorporates a variety of market-specific factors and
errors of cross-sectional portfolio sorts indicate that
peculiarities (Sarr and Lybek 2002). Therefore, in
the higher MSEs for the historical illiquidity bench-
addition to examining historical illiquidity, we con-
mark describe a general pattern and are not driven
sider a comprehensive set of bond characteristics
by high forecast errors for only a few bonds with
and exploit their information content using machine
extreme characteristics.
learning models. We apply both relatively simple lin-
ear models (with and without penalty terms for multi- In addition to analyzing the differences in level fore-
ple predictors) and more complex models that casts of illiquidity, we assess the economic value of
capture patterns of nonlinear and interactive effects illiquidity forecasts on the basis of a portfolio forma-
in the relationship between predictor variables and tion exercise, that is, by trading bonds sorted into
expected bond illiquidity (such as regression tress portfolios based on realized and expected illiquidity.
and neural networks).2 Another limitation in earlier Compared to the historical illiquidity–based bench-
work is that it considers mostly the bid–ask spread mark, machine learning forecast models are better at
even if, in many cases, it is not a good representation disentangling more liquid from less liquid bonds.
of a bond trade’s realistic costs of execution. This is Moreover, following Amihud and Mendelson (1986),
because the tradability of a bond itself and the mar- investors should require higher expected returns for
ket impact of the trade also have a substantial more illiquid bonds to compensate for higher trading
effect on investment performance. In our analysis, expenses. Confirming this notion, we find that

104
 Predicting Corporate Bond Illiquidity via Machine Learning

prediction models using machine learning techniques returns typically leads to the risk of overfitting
generate a higher illiquidity premium in the cross- machine learning models. In contrast, bond liquidity,
section of bond returns than the historical illiquidity our variable of interest, is highly persistent, that is,
benchmark model. We highlight the economic value past relations are more likely to continue to hold in
added in numerical examples and showcase that the future, resulting in a higher signal-to-noise ratio.
even small improvements in illiquidity estimates can Therefore, machine learning methods should work at
result in large transaction cost savings, either directly least as well or maybe even better for predicting
in terms of a lower average bid–ask spread or indi- future bond illiquidity than they do for predicting
rectly in terms of a lower average price impact. future stock and bond returns. Second, from a trad-
ing and execution perspective, a better representa-
Furthermore, using relative variable importance met- tion of the expected illiquidity dimension in bond
rics, we document that the historical illiquidity–based trading should provide economic value added for
predictor is most important. This is because realized
investors. Given the speed and complexity of bond
bond illiquidity is highly persistent and has long-
trading, machine learning methods can help to exploit
memory properties. Among the remaining variables,
alpha signals even after accounting for transaction
fundamental and risk- as well as return-based covari-
costs such that bond investors will embrace machine
ates are the most important predictors (in that order).
learning methods as an essential part of their trading
Macroeconomic indicators seem much less informa-
practices in the future. Our results suggest that more
tive for future illiquidity. However, variable impor-
complex machine learning models tend to be more
tance itself is also time-varying, and even predictors
powerful. However, because the implementation of
that are unconditionally less informative play impor-
these methods requires significant resources and
tant roles at times. Consequently, it is important to
skills, the choice of a specific type of prediction
apply prediction models that are able to accommo-
model will depend on how practitioners use illiquidity
date the time-varying nature of illiquidity indicators.
forecasts in their bond investment and trading
By way of an example, we address this “black box”
decisions.
characteristic and illustrate how machine learning
estimators for bond illiquidity generate value for
investors. In particular, we visualize the combined
Literature Review
effect of duration and rating on a bond’s illiquidity
Previous literature documents that a bond’s illiquidity
estimate, which confirms that a large part of the pre-
evolves throughout its lifetime (Warga 1992; Hong
diction outperformance of the more complex
and Warga 2000; Hotchkiss and Jostova 2017), sug-
machine learning models is due to their ability to
gesting that dynamic estimation methods, such as the
exploit nonlinear and interactive patterns.
machine learning models we use, may be promising
Based on empirical evidence from predicting stock candidates for predicting bond illiquidity. Moreover,
returns, several other recent papers take a more time-varying bond characteristics, such as size (Bao,
skeptical position on the use of machine learning in Pan, and Wang 2011; Jankowitsch, Nashikkar, and
asset management applications. For example, Subrahmanyam 2011) and risk (Mahanti et al. 2008;
Avramov, Cheng, and Metzker (2022) conclude that Hotchkiss and Jostova 2017), impact expected bond
machine learning signals extract a large part of their illiquidity. Therefore, applying machine learning tech-
profitability from difficult-to-arbitrage stocks (dis- niques, which adaptively incorporate these features
tressed stocks and microcaps) and during high limits- along with their nonlinearities and interactions,
to-arbitrage market states (high–market volatility should be valuable for predicting bond illiquidity.
periods). Moreover, they document that machine
learning–based performance will be even lower Empirical evidence indicates that machine learning
because of high turnover and trading costs. Similarly, methods are able to outperform established
Leung et al. (2021) show that the extent to which approaches in various prediction tasks. Examples
the statistical advantage of machine learning models include forecasting stock returns (Gu, Kelly, and Xiu
can be translated into economic gains depends on 2020; Freyberger, Neuhierl, and Weber 2020), pre-
dicting bond risk premiums (Bianchi, Bu €chner, and
the ability to take risk and implement trades
efficiently. Tamoni 2021; Bali et al. 2022), and modeling stock
market betas (Drobetz et al. 2024). Realized bond illi-
Our work contributes to this strand of more critical quidity, however, is much less noisy than realized
work in the machine learning literature in two impor- stock and bond returns. Compared to return series
tant ways. First, the low signal-to-noise ratio in stock but similar to beta variation, illiquidity is highly

Volume 80, Number 3 105

Financial Analysts Journal | A Publication of CFA Institute

persistent over time. Given a higher signal-to-noise intraday transaction records for the US corporate
ratio, estimating future corporate bond illiquidity bond market reported in the enhanced version of
should provide a sensible use case for the application TRACE for the sample period from July 2002 to
of machine learning techniques. December 2020. The TRACE dataset comprises the
most comprehensive information on US corporate
The study most closely related to our work is from bond transactions, with intraday observations on
Reichenbacher, Schuster, and Uhrig-Homburg (2020). price, transaction volume, and buy and sell
They apply linear models to predict future corporate indicators.3 In addition, bond characteristics (issue
bond bid–ask spreads, which they use as their proxy information) such as bond type, offering and
for liquidity although it ignores the potentially large maturity dates, coupon specifications, outstanding
market impact of a trade. While these authors also amount, rating, and issuer information come from
use a large set of predictor variables and analyze Mergent FISD.4
their importance, they do not explore patterns of
nonlinear and interactive effects in the relation To clean the TRACE dataset, we use Dick-Nielsen’s
between predictor variables and bond illiquidity esti- (2009, 2014) procedure to remove duplicate, can-
mates. In our own analysis, we extend their insightful celled, and corrected entries. Following Bali,
work in several directions. Most important, (1) we Subrahmanyam, and Wen (2021), we omit bonds
compare the predictive performance of machine from the sample that (1) are not listed or traded in
learning estimators to that of the commonly used his- the US public market; (2) are backed with a guaran-
torical illiquidity benchmark, (2) we analyze how tee or linked to an asset; (3) have special features
machine learning models outperform by assessing (perpetuals, convertible and puttable bonds, or float-
forecast errors of cross-sectional portfolio sorts, (3) ing coupon rates); or (4) have less than one year to
we use a comprehensive set of liquidity measures maturity or are defaulted. For the intraday records,
that also captures a bond trade’s market impact, (4) we eliminate transactions that (5) are labeled as
we assess the economic value added of machine when-issued or locked-in or have special sales condi-
learning–based estimators, and (5) we scrutinize the tions; (6) have more than a three-day settlement; and
importance of nonlinear and interactive effects in (7) have a volume less than $10,000 or a price less
establishing illiquidity predictions. than $5.

Our paper is related to recent studies that use Based on the intraday bond transaction records, we
machine learning in various fixed-income applications. aggregate our database on a monthly basis and con-
For example, Fedenia, Nam, and Ronen (2021) show struct three distinct illiquidity measures that capture
that random forest algorithms can be used to different aspects of illiquidity. All variables used in
uncover a better trade signing model in the corporate our empirical analyses are described in Table 1. First,
bond market, that is, to determine whether a trade is we consider the transaction volume (t volumet ), which
buyer- or seller-initiated, which helps bond traders to is related to the capacity of actually trading the
better understand market dynamics and price behav- respective bond:
ior. Cherief et al. (2022) apply random forests and X
Nt
gradient-boosted regression trees to capture nonli- t volumet ¼ Qd , (1)
nearities and interactions between traditional risk d¼1

factors in the credit space. Their model outperforms where Qd is the dollar trading volume on day d, and
linear pricing models in forecasting credit excess Nt is the number of trading days with positive-trading
returns. Kaufmann, Messow, and Vogt (2021) use volume in each month t: Second, following Hong and
gradient-boosted regression trees to model the Warga (2000) and Chakravarty and Sarkar (2003), we
equity momentum factor (in addition to classical compute the difference between the average cus-
bond market factors such as size and illiquidity) in tomer buy and the average customer sell price on
the corporate bond market. each day within a given month t (t spreadt ) to quan-
tify transaction costs:

Data 1 XNt PBuy

d − Pd
Sell
t spreadt ¼ , (2)
Nt d¼1
Following Bessembinder, Maxwell, and Venkataraman 0:5  PBuy
d þ PSell
d
(2006), who emphasize the importance of using
Buy=Sell
Trade Reporting and Compliance Engine (TRACE) where Pd is the average price of customer buy/
transaction data, our empirical analysis is based on sell trades on day d: Third, we use Amihud’s (2002)

106
 Predicting Corporate Bond Illiquidity via Machine Learning

Table 1. Variable Descriptions and Definitions

# Predictor Description Definition

Predictors based on historical illiquidity

1a t_volume_hist Historical transaction volume Log historical transaction volume (computed as
the average over the last 12 months)
1b t_spread_hist Historical transaction cost Log historical transaction cost (computed as the
average over the last 12 months)
1c amihud_hist Historical price impact Log historical price impact (computed as the
average over the last 12 months)
Fundamental predictors
2 age Age Bond age since first issuance, measured in
number of years
3 size Size Log bond amount outstanding
4 rating Rating Numerical bond rating calculated as described
in Bali et al. (2020), measured from 1 (good)
to 22 (bad)
5 mat Maturity Time to maturity, measured in number of years
6 yield Yield Bond yield to maturity
7 dur Duration Bond price sensitivity to interest rate changes,
measured in number of years
Return-based predictors
8 rev Short-term reversal Excess return in the prior month
9 mom Momentum Excess return from month −12 to month −1
10 ltr Long-term reversal Excess return from month −36 to month −1
11 vol Volatility Standard deviation, estimated from monthly
returns over the last three years
12 skew Skewness Skewness, estimated from monthly returns
over the last three years
13 kurt Kurtosis Kurtosis, estimated from monthly returns over
the last three years
Risk-based predictors
14 var10 Value at Risk Value at risk at the 10% level, measured as
fourth lowest monthly return observation
over the last three years
15 es10 Expected Shortfall Expected shortfall at the 10% level, measured
as average of the four lowest monthly return
observations over the last three years
16 beta Systematic risk Bond market beta, measured as regression
coefficient from the time-series ordinary
least squares regression of monthly bond
excess returns on market excess returns over
the last three years
17 idio Idiosyncratic risk Idiosyncratic bond risk, measured as mean
squared error of the residuals from the
time-series ordinary least squares regression of
monthly bond excess returns on market excess
returns over the last three years
Macroeconomic indicators
18 dfy Default spread Yield differential between Moody’s Baa- and
Aaa-rated corporate bonds

Notes: This table shows descriptions and definitions for each of the 18 bond illiquidity predictors used in the empirical analysis.
The baseline sample includes intraday transaction records for the US corporate bond market reported in the enhanced version
of TRACE for the sample period from July 2004 to November 2020.

Volume 80, Number 3 107

Financial Analysts Journal | A Publication of CFA Institute

Table 2. Cross-Sectional and Time-Series Correlations between Illiquidity Measures

Panel A: Cross-sectional Panel B: Time-series
t_volume amihud t_spread t_volume amihud t_spread

t_volume 1.00 −0.63 −0.09 1.00 −0.41 0.06

amihud 1.00 0.42 1.00 0.25
t_spread 1.00 1.00

Notes: This table shows cross-sectional and time-series correlations among the three realized bond illiquidity measures, t_volume,
amihud, and t_spread. Panel A (cross-sectional) contains the time-series averages of monthly cross-sectional correlations, and Panel
B (time-series) the cross-sectional averages of time-series correlations. The baseline sample includes intraday transaction records
for the US corporate bond market reported in the enhanced version of TRACE for the sample period from July 2004 to
November 2020.

Table 3. Transition Probabilities

Decile
1 2 3 4 5 6 7 8 9 10

t_volume Prob (no transition), % 45.75 28.70 22.26 19.16 17.39 17.00 17.62 20.70 29.22 62.47
t value 88.54 82.57 59.91 53.58 46.53 40.05 48.96 54.94 76.65 168.25
amihud Prob (no transition), % 37.03 19.47 15.75 13.79 13.61 13.31 13.85 14.86 17.65 25.06
t value 39.89 37.68 35.76 26.75 28.13 23.42 25.85 30.52 44.71 51.87
t_spread Prob (no transition), % 27.54 23.40 19.10 17.10 16.14 15.90 16.86 18.88 23.91 43.15
t value 60.37 59.54 50.36 40.81 36.50 36.70 39.11 42.52 63.83 78.33

Notes: Based on monthly sortings of bonds into illiquidity deciles, this table shows average transition probabilities (together with one-
sided t-statistics) for all three realized bond illiquidity measures (t_volume, amihud, and t_spread). Only the diagonal elements of the full
transition matrix are shown, that is, the average probabilities to remain in the same illiquidity decile in the subsequent month (“no-tran-
sition” probabilities). The baseline sample includes intraday transaction records for the US corporate bond market reported in the
enhanced version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

measure of illiquidity (amihudt ), which captures the for the three bond illiquidity measures. To keep the
aggregate price impact in each month t : table tractable, we only show the diagonal elements
of the full transition matrix, that is, the average
1X Nt
jrd j probabilities to remain in the same illiquidity decile
amihudt ¼  106 , (3)
Nt d¼1 Qd in the subsequent month. These “no-transition”
probabilities all exceed 10%, confirming that bond
P
where rd ¼ Pi, it−1
,t
− 1 is a bond’s price return on day d.5 illiquidity is persistent (Chordia, Sarkar, and
Subrahmanyam, 2005; Acharya, Amihud, and
Table 2 presents cross-sectional and time-series Bharath, 2013) and suggesting that lagged historical
correlations of these three realized bond illiquidity illiquidity will be a particularly important predictor
measures. Panel A (cross-sectional) contains the variable for expected illiquidity.
time-series averages of monthly cross-sectional cor-
relations, and Panel B (time-series) the cross-sectional In addition to a predictor based on realized illiquidity
averages of time-series correlations. The correlations over the past year that captures the time-series
(in absolute values) are far from perfect and range dynamics in illiquidity, we select from Bali,
widely between −0.63 and þ0.25, confirming that Subrahmanyam, and Wen (2021) a comprehensive
our three measures capture different aspects of set of 18 forecasting variables, which are described
illiquidity. in Table 1. These variables capture basic return and
risk characteristics of bonds.6 While these variables
Based on monthly sortings of bonds into illiquidity describe the characteristics of bonds in general and
deciles, Table 3 presents the average transition are somehow natural candidates for our forecasting
probabilities (together with one-sided t-statistics) task, they need not even be the best predictors for

108
 Predicting Corporate Bond Illiquidity via Machine Learning

expected bond illiquidity. Our set of predictor varia- rather than the marginal effects of each single pre-
bles includes six fundamental predictors based on the dictor. Machine learning methods are suitable for
characteristics of bonds (age, size, rating, maturity, solving the multicollinearity problem either by nature
duration, and yield). In addition, it contains 10 techni- (tree-based models) or by applying different types of
cal indicators based on the historical bond return dis- regularization, for example, a lasso-based penalization
tribution, relating to return characteristics (short-term of the weights (neural networks).
reversal, momentum, long-term reversal, volatility,
skewness, and kurtosis) and risk characteristics
(value-at-risk, expected shortfall, systematic risk, and Forecast Models
idiosyncratic risk). Technical indicators are computed
based on monthly excess bond returns:
General Approach. Our objective is to examine
  whether machine learning methods outperform the
Pi, t þ AIi, t þ Ci, t historical illiquidity benchmark model, that is, the
ri, t ¼ Ri, t − rf, t ¼ − 1 − rf , t , (4)
Pi, t−1 þ AIi, t−1 naïve rolling-window approach, in terms of predictive
where Pi, t is the transaction price, AIi, t is the accrued performance and, if yes, why. We are particularly
interest, and Ci, t is the coupon payment, if any, of interested in examining whether (1) incorporating our
bond i in month t: Based on the TRACE records, we large set of bond characteristics as predictors and (2)
first calculate the daily clean price as the transaction allowing for nonlinearity and interactions in the rela-
volume-weighted average of intraday prices to mini- tionship between these predictors and future
mize the effect of bid–ask spreads in prices, follow- (expected) illiquidity can add incremental predictive
ing Bessembinder et al. (2009), and then convert the power. We run a horse race between the benchmark
bond prices from daily to monthly frequency by model that uses historical illiquidity (the average illi-
keeping the price at the end of a given month t: rf, t quidity over the last 12 months) and linear as well as
is the risk-free rate proxied by the US Treasury bill nonlinear machine learning–based prediction models
rate. If necessary, the value-weighted portfolio of all that exploit additional cross-sectional information,
bonds serves as the proxy for the market portfolio. comparing their performance from both a statistical
Finally, our analysis includes the default spread as and an economic perspective. In addition, we analyze
the only macroeconomic covariate. the characteristics and functioning scheme of the
machine learning techniques that help explain their
We only include a bond in our empirical analysis for superior predictive performance.
month t if the illiquidity measure under investigation
is available and the bond provides complete informa- Following the approach used for estimating market
tion on all predictor variables, that is, there are no betas in Cosemans et al. (2016) and Drobetz et al.
missing values. In every month, we require at least (2024), the estimation setting in our empirical tests is
100 bonds to be included in the cross-section. This as follows: Out-of-sample illiquidity estimates are
limits our sample period to July 2004 through obtained at the bond level and on a monthly basis,
November 2020. The average monthly cross-section following an iterative procedure. In the first iteration
consists of 4,330 bonds for t_volume, 3,556 bonds step, we use data up to the end of month t and
for t_spread, and 4,328 bonds for amihud. obtain forecasts for each bond i’s average monthly illi-
quidity during the out-of-sample forecast period
As in Cosemans et al. (2016), we winsorize outliers in (from the beginning of month t þ 1 to the end of
both the illiquidity measures and all predictors month t þ k), denoted as IFi, tþkjt (or abbreviated IFi, t ).
(except the default spread) to the 1st and 99th per- We set k equal to 12, focusing on a one-year fore-
centile values of their cross-sectional distributions. cast horizon.7 In the next iteration step, we use data
Moreover, we correct for skewness in distributions up to the end of month t þ 1 and obtain forecasts of
by logarithmically transforming the three illiquidity bond-level illiquidity during the subsequent out-of-
measures and some of the predictor variables (see sample forecast period (from the beginning of month
Table 1). Some predictors are constructed similarly— t þ 1 þ 1 to the end of month t þ 1 þ k). By iterating
for example, value-at-risk and expected shortfall—or through the entire sample, we obtain time-series of
incorporate similar information—for example, matu- overlapping annual out-of-sample illiquidity predic-
rity and duration—which leads to relatively high cor- tions, which we compare to realized illiquidity.
relations. However, according to Lewellen (2015),
multicollinearity is not a main concern in our setup Next, we introduce the different models used to pre-
because we are mostly interested in the overall pre- dict future bond illiquidity. Online Supplemental
dictive power of machine learning–based models Appendix A provides details. While they differ in

Volume 80, Number 3 109

Financial Analysts Journal | A Publication of CFA Institute

their overall approach and complexity, all models aim of the conditional expectations g ðzi, t Þ are flexible
to minimize the forecast error of level predictions, and family-specific. Approximation functions gð:Þ can
defined as the MSE at the end of each month t : be linear or nonlinear. Moreover, they can be para-
X
Nt metric, with gðzi, t , hÞ, where h is the set of true
MSEtþkjt ¼ ðIRi, tþk − IFi, tþkjt Þ2 , with k ¼ 12, (5) parameters, or nonparametric, with gðzi, t Þ:
i¼1

where IRi, tþk is bond i’s realized average monthly illi- A general problem is that machine learning methods
quidity during the out-of-sample period (i.e., from the are prone to overfitting, which is why we must con-
beginning of month t þ 1 to the end of month t þ k), trol for the degree of model complexity by tuning
and Nt is the number of bonds at the end of the relevant hyperparameters. To avoid overfitting
month t: and maximize out-of-sample predictive power, the
hyperparameters should not be preset, but rather
Benchmark Estimator. Most academic papers must be determined adaptively from the sample data.
that focus on the bond market use a bond’s We follow Gu, Kelly, and Xiu’s (2020) time-series
historical illiquidity as a naïve prediction for future cross-validation approach to fit the machine learn-
illiquidity (Bao, Pan, and Wang 2011; Friewald, ing–based forecast models so that they produce reli-
Jankowitsch, and Subrahmanyam 2012; Dick- able out-of-sample predictive performance. Online
Nielsen, Feldhu €tter, and Lando 2012; Bongaerts, de Supplemental Appendix A provides details on how
Jong, and Driessen 2017). Given the high persis- we split the sample into three subsamples: a training
tence in realized bond illiquidity (see Table 2), we sample, a validation sample, and a test sample. We
implement this naïve estimator in all our empirical obtain our first illiquidity estimates in June 2011,
tests. Since we focus on a one-year forecast hori- using six years of data for training and validation
zon, we use the average monthly illiquidity over the (2004:07–2009:06 and 2009:07–2010:06, respec-
last 12 months (t volume hist, t spread hist, and tively), which we then compare to the bonds’ realized
amihud hist) as our benchmark, thereby increasing illiquidity over the next year.8 This approach ensures
the signal-to-noise ratio relative to the current- that our test sample is truly out-of-sample, enabling
month illiquidity. us to evaluate a model’s out-of-sample predictive
power. In total, we exploit eight years and six months
Machine Learning Estimators. Rather than
of data for testing (up to the end 2019:11).
simply averaging historical illiquidity measures,
machine learning techniques focus explicitly on the As already explained, we consider a set of 18 predic-
objective of forecasting corporate bond illiquidity. tor variables (predictors based on historical illiquidity,
Realized illiquidity enters our regressive framework fundamental predictors, return-based predictors, risk-
as the dependent variable, while historical illiquidity, based predictors, and macroeconomic indicators; see
a set of bond characteristics, and macroeconomic Table 1 for more details) to fit the machine learning
indicators serve as predictors. We adapt the additive
techniques. We test three different forecast model
prediction error model from Gu, Kelly, and Xiu (2020)
families, which differ in their overall approach and
to describe a bond’s illiquidity:
complexity. Online Supplemental Appendix A pro-
IRi, tþk ¼ Et ðIRi, tþk Þ þ ei, tþk , (6) vides more details on these techniques and how we
implement them.
where IRi, tþk is bond i’s realized illiquidity over the
one-year forecast horizon starting at the beginning of The first model family consists of linear regressions,
month t þ 1: Expected illiquidity is estimated as a for which we use the training sample to run pooled
function of multiple predictor variables and described ordinary least squares regressions of future realized
by the “true” model g ðzi, t Þ, where zi, t represents the illiquidity IRi, tþk on the set of 18 predictors. We either
P-dimensional set of predictors: use the ordinary least squares loss function (lm) or
 
modify it by incorporating a penalty term, that is, we
Et IRi, tþk ¼ g ðzi, t Þ: (7)
apply an elastic net penalization (elanet). The latter is
Although our machine learning–based forecast mod- the most common machine learning technique to
els belong to different families (linear regressions, overcome the overfitting problem in high-dimensional
tree-based models, and neural networks), they are all regressions, for example, when the number of predic-
designed to approximate the true forecast model by tors becomes large relative to the number of obser-
minimizing the out-of-sample MSE. Approximations vations. If not explicitly included as predetermined

110
 Predicting Corporate Bond Illiquidity via Machine Learning

terms, pooled regressions (simple or penalized mod- cross-sectional averages of expected illiquidity, (2)
els) cannot capture nonlinear or interactive effects. cross-sectional standard deviations, and (3) cross-
sectional minimum, median, and maximum values.
The second model family consists of tree-based mod- Following Pastor and Stambaugh (1999), we report
els, for which we use random forests (rf) and gradi- the implied cross-sectional standard deviation of true
ent-boosted regression trees (gbrt), the most  1=2
common models within this category. Finally, the c ðIR Þ ¼ VarðIF Þ − Var
illiquidity, Std dR , which helps
Ii
third model family comprises neural networks (nn_1–
nn_5), for which we consider specifications with up to measure an illiquidity forecast’s precision. The
to five hidden layers and 32 neurons.9 Both tree- minuend VarðIF Þ is the time-series average of
based models and neural networks incorporate monthly cross-sectional variances, and the subtra-
nonlinearities and multiway interactions inherently, d R denotes the cross-sectional average of
hend Var Ii
without the need to add new predictors to capture
bonds’ sampling variance. Small gaps between
these effects.
observed and implied standard deviations imply small
estimation errors, indicating measurement of true illi-
quidity with high precision. Panel B focuses on time-
Empirical Results series properties, presenting the cross-sectional
Having introduced the benchmark and machine learn-
means of (1) time-series averages of estimated illi-
ing–based estimation approaches, we now apply quidity; (2) time-series standard deviations; (3) time-
these models to forecast out-of-sample bond illiquid- series minimum, median, and maximum values; and
ity. We focus on the amihud measure in presenting (4) first-order autocorrelations.
and discussing the empirical results going forward
because what matters most to investors is the actual The cross-sectional and time-series means for each
price impact their trades will have. The return pre- estimation approach are close to those for realized
mium associated with this illiquidity measure is gen- illiquidity,11 while the cross-sectional and time-series
erally considered an illiquidity risk premium that dispersions vary across the models. Standard devia-
compensates for price impact or transaction costs. tions (SDs) are greatest for the hist model, which
Our results are qualitatively similar for the alternative uses only time-series information based on a bond’s
t_volume and t_spread measures. Supplementary historical illiquidity. This restriction leads to extreme
Appendix C presents our main results using these and highly volatile illiquidity estimates. In contrast,
two bond illiquidity measures together with other incorporating cross-sectional information about a
robustness tests. bond’s characteristics, its return-risk profile, and mac-
roeconomic indicators reduces the cross-sectional
We start with studying the models’ ability to predict and time-series standard deviations in expected illi-
bond illiquidity from a statistical perspective. Our quidity notably. Since this reduction in volatility is
focus is on the question whether machine learning– similar for all machine learning models, it seems to be
based level forecasts of illiquidity outperform the his- the inclusion of slow-moving bond characteristics as
torical illiquidity benchmark. We assess the cross- predictors in the additive prediction error model
sectional and time-series properties of our models’ rather than the ability of the more complex models
prediction performance, particularly comparing the to capture nonlinearity and interactions that results
resulting forecast errors. We also investigate the in less extreme and less volatile estimates. In other
underlying causes of differences in predictive perfor- words, the time variation in bond characteristics is
mance by analyzing the forecast errors of cross- able to pick up long-run movements in illiquidity.
sectional portfolio sorts. Moreover, we evaluate
whether differences in statistical predictive perfor- The observed cross-sectional SD of illiquidity fore-
mance translate into economic gains in a portfolio casts in Panel A is most informative for the assess-
formation exercise. ment of a model’s precision when comparing it to
the implied cross-sectional standard deviation of true
Cross-Sectional and Time-Series illiquidity (Impl. SD). This comparison reveals that
Properties of Illiquidity Estimates. To begin, true illiquidity is measured with the lowest precision
we investigate the properties of illiquidity predictions (implying larger gaps between observed and implied
obtained from the different forecast models.10 Panel SDs) by the historical illiquidity–based benchmark
A in Table 4 focuses on the cross-sectional proper- model and with the highest precision (implying
ties, presenting the time-series means of monthly (1) smaller gaps) by the machine learning–based models.

Volume 80, Number 3 111

Financial Analysts Journal | A Publication of CFA Institute

Table 4. Cross-Sectional and Time-Series Properties of Illiquidity Estimates

Panel A: Cross-sectional Panel B: Time-series
Mean SD Min Median Max Impl. SD Mean SD Min Median Max Autocorr.

hist −4.77 0.59 −6.97 −4.67 −3.80 0.49 −4.79 0.29 −5.45 −4.77 −4.30 0.91
lm −4.78 0.43 −6.45 −4.72 −3.73 0.39 −4.80 0.18 −5.22 −4.79 −4.49 0.91
elanet −4.77 0.42 −6.39 −4.71 −3.77 0.38 −4.79 0.17 −5.17 −4.78 −4.49 0.91
rf −4.78 0.42 −6.72 −4.67 −4.09 0.39 −4.80 0.16 −5.16 −4.78 −4.54 0.90
gbrt −4.77 0.43 −6.84 −4.66 −3.93 0.39 −4.79 0.17 −5.19 −4.77 −4.50 0.88
nn_1 −4.79 0.43 −6.84 −4.69 −3.90 0.38 −4.80 0.18 −5.23 −4.78 −4.50 0.92

Notes: Properties of out-of-sample bond illiquidity estimates (the average monthly amihud measure) are obtained from the differ-
ent forecast models (hist, lm, elanet, rf, gbrt, and nn_1). Panel A focuses on cross-sectional properties, presenting time-series means
of (1) the value-weighted cross-sectional average of estimated bond liquidity, (2) the cross-sectional standard deviation, and (3)
the cross-sectional minimum, median, and maximum value. Following the procedure outlined in Paster and Stambough (1999), it
 1=2
c ðIR Þ ¼ VarðIF Þ − Var
also reports the implied cross-sectional standard deviation of true bond illiquidity, that is, Std dR : Panel B
Ii

focuses on time-series properties, presenting value-weighted cross-sectional means of (1) the time-series average of bond illiquid-
ity’ (2) the time-series standard deviation; (3) the time-series minimum, median, and maximum value; and (4) the first-order auto-
correlation. Following Becker et al. (2021), firms with fewer than 50 bond illiquidity estimates are omitted for the summary
statistics in Panel B. The baseline sample includes intraday transaction records for the US corporate bond market reported in the
enhanced version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

For example, the difference between SD and Impl. historical illiquidity–based benchmark by more than
SD is 0.1 for the hist model and only 0.05 for the 23%. We conclude that these models’ ability to cap-
nn_1 model. Finally, although they incorporate slow- ture nonlinearity and interactions further enhances
moving bond characteristics as predictors, the aver- the quality of illiquidity predictions by reducing the
age time-series autocorrelations of machine learning– forecast error of level predictions.
based models in Panel B are similar to that of the
Since, by construction, these figures reflect a forecast
historical illiquidity benchmark model (all
model’s average predictive performance, we next
around 0.90).
investigate the forecast errors over time. Panel B of
Average Forecast Errors and Forecast Table 5 reports the fraction of months during the
out-of-sample period for which the column model (1)
Errors over Time. Next, we examine the statisti-
is in the Hansen, Lunde, and Nason (2011) model
cal predictive performance of the different forecast
confidence set (MCS) and (2) is significantly better
models by comparing their forecast errors. Panel A of
than the row model in a pairwise comparison (accord-
Table 5 reports the time-series means of monthly
ing to Diebold and Mariano (1995) test [DM test]
MSEs (based on a one-year forecast horizon), calcu-
statistics). The MCS approach incorporates an adjust-
lated as specified in Equation (5). Exploiting only
ment for multiple testing and is designed to include
bond-level time-series information, the estimates
the best forecast model(s) based on a certain confi-
based on historical illiquidity generate sizable fore-
dence level.12 The DM test of equal predictive ability
cast errors (0.192 in the hist model). Incorporating
inspects pairwise differences in bond-level squared
cross-sectional information reduces the average MSE forecast errors (SEs):
noticeably. Linear regressions (both simple and penal-
ized, with MSEs of 0.160 and 0.157, respectively) SEi, tþkjt ¼ ðIRi, tþk − IFi, tþkjt Þ2 , with k ¼ 12: (8)
reduce the average forecast error relative to the hist
The DM test statistic in month t for comparing model
model by around 18%. Inspecting nonlinear machine
d
learning methods, we find that tree-based models j with a competing model i is DMij, t ¼ r^ ij, t , where
d ij, t
and neural networks reduce the average forecast
dij, t ¼ SEi, tþkjt − SEj, tþkjt is the difference in SEs, d ij, t ¼
error relative to linear regressions even further (with PNt
average MSEs of 0.145, 0.144, and 0.147 for the rf, i¼1 dij, t is the cross-sectional average of these dif-
gbrt, and nn_1 model, respectively). Tree-based mod- ferences, and r ^ d denotes the heteroscedasticity-
ij, t

els and neural networks perform similarly well, and autocorrelation-consistent standard error of dij, t :
decreasing the average forecast error relative to the We use the Newey and West’s (1987) estimator with

112
 Predicting Corporate Bond Illiquidity via Machine Learning

Table 5. Forecast Errors

Forecast model
hist lm elanet rf gbrt nn_1

Panel A: Average forecast errors

MSE 0.192 0.160 0.157 0.145 0.144 0.147
Panel B: Forecast errors over time
In MCS 0.00 2.94 7.84 51.96 74.51 50.98
vs. hist 92.16 98.04 94.12 99.02 94.12
vs. lm 1.96 40.20 88.24 86.27 87.25
vs. elanet 0.00 19.61 83.33 79.41 77.45
vs. rf 0.00 2.94 6.86 46.08 22.55
vs. gbrt 0.00 0.00 2.94 14.71 17.65
vs. nn_1 1.96 5.88 10.78 30.39 46.08
T 102 102 102 102 102 102

Notes: This table presents differences in forecast errors for the amihud illiquidity measure produced by the forecast models (hist,
lm, elanet, rf, gbrt, and nn_1). Panel A reports the time-series means for monthly value-weighted mean-squared errors (MSEs), that
P t
is, MSEtþkjt ¼ Ni¼1 wi, t ðIRi, tþk − IFi, tþkjt Þ2 , with k ¼ 12, where Nt is the number of bonds in the sample at the end of month t: Panel B
reports the fraction of months during the out-of-sample period for which the column model is (1) in the Hansen, Lunde, and
Nason (2011) model confidence set (MCS) and (2) significantly better than the row model in a pairwise comparison (according to
Diebold and Mariano (1995) test [DM test] statistics). The DM tests of equal predictive ability inspect differences in stock-level
squared forecast errors (SEs), that is, SEi, tþkjt ¼ ðIRi, tþk − IFi, tþkjt Þ2 , with k ¼ 12: The DM test statistic in month t for comparing the
d
model under investigation j with a competing model i is DMij, t ¼ r^ ij, t , where dij, t ¼ SEi, tþkjt − SEj, tþkjt is the difference in SEs, d ij, t ¼
d ij, t
PNt
i¼1 dij, t is the cross-sectional average of these differences, and r
^ d ij, t is the Newey and West (1987) estimator with four lags to
account for possible heteroskedasticity and autocorrelation. Positive signs of DMij, t indicate superior predictive performance of model j
relative to model i in month t, that is, that model j yields, on average, lower forecast errors than model i: All statistical tests are based
on the 5% significance level. The baseline sample includes intraday transaction records for the US corporate bond market reported in
the enhanced version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

four lags to compute standard errors and follow the methods, in a statistical sense, provide higher quality
convention that positive signs of DMij, t indicate supe- level forecasts of bond illiquidity.
rior predictive performance of model j relative to
model i in month t, that is, that model j yields, on Supporting this finding, the results from the monthly
average, lower forecast errors than model i:13 DM tests overwhelmingly show that all machine
learning methods dominate the historical illiquidity–
We observe that the historical illiquidity benchmark based model in pairwise comparisons, with fractions
model is in the MCS of the best forecast models in ranging from 92.16% for the lm model to 99.02% for
none of the 102 months during our sample period. In the gbrt model of all sample months. This suggests
other words, for every single month, we can reject the that machine learning models are superior to the
null hypothesis that the hist benchmark model gener- benchmark model in different states of the world,
ates the best illiquidity forecasts. The percentages of that is, in both “normal” market phases as well as
months for which linear regressions (simple and penal- phases of market turmoil.14
ized, with 2.94% and 7.84%, respectively) are in the
MCS of the best models are very low as well. In sharp contrast, the hist model rarely yields signifi-
Regression trees and neural networks are in the MCS cantly lower MSEs than the machine learning–based
of the most accurate forecast models considerably approaches (as indicated by the low fractions of
more often, ranging from 50.98% of the months for months, ranging between 0.00% vs. the elanet, rf,
the nn_1 model and 74.51% of the months for the gbrt and gbrt models and 1.96% vs. the lm and nn_1 mod-
model. Put differently, we must reject the null hypothe- els). Moreover, tree-based models and neural net-
sis that the nn_1 model and the gbrt model are among works dominate linear regressions (both linear and
the best forecast models in only about 41% and 25% penalized) in at least 77.45% of the months, while
of months, respectively. Taken together, these findings their linear counterparts yield a significantly lower
strongly suggest that the nonlinear machine learning MSE in only 10.78% of the months or even less.

Volume 80, Number 3 113

Financial Analysts Journal | A Publication of CFA Institute

Overall, the results indicate outperformance of non- the forecast errors relative to the hist (lefthand col-
linear machine learning models over the historical illi- umn) and lm (righthand column) models for nearly all
quidity benchmark and linear regressions.15 decile portfolios. This is indicated by percentage dif-
Comparing the machine learning techniques, with the ferences larger than zero (the squares above the
aim of generating low forecast errors, the gbrt model dashed line), implying that the nn_1 model delivers
performs the best. Gradient-boosted regression trees more accurate illiquidity predictions. The figure fur-
exhibit the largest MCS fraction and surpass random ther emphasizes that the higher average MSEs for
forests as well as neural networks in illiquidity predic- the historical illiquidity–based approach and linear
tion significantly more often than they are dominated regressions (see Panel A of Table 5) obey more gen-
by them. Moreover, the random forest model (rf), our eral patterns and are not driven by high forecast
second tree-based model, also seems to be slightly errors for only a few bonds with specific characteris-
superior to the simplest neural network model tics. Compared to the two benchmark models, the
(nn_1).16 reduction in forecast errors when using neural net-
works are strongest for extreme decile portfolios
Forecast Errors of Cross-Sectional (which are more difficult to predict), both in absolute
Portfolio Sorts. In additional analysis, we exam- and relative terms. Because this pattern is apparent
ine cross-sectional differences in the performance of for the comparison with both the historical illiquid-
machine learning models relative to the historical illi- ity–based approach and linear regressions, we attri-
quidity benchmark. We attempt to identify types of bute the reduction in forecast errors to two effects:
bonds, for example, larger vs. smaller bonds, for (1) the inclusion of slow-moving bond characteristics
which the differences in forecast errors across illi- as predictors (in both the lm and the nn_1 model)
quidity estimators are most pronounced. We sort all and (2) the nn_1 model’s ability to capture nonlinear-
bonds into decile portfolios based on their character- ity and interactions.
istics, that is, historical illiquidity (amihud_hist), size
(size), rating (rating), systematic risk (beta), and idio- Characteristics of Expected Illiquidity-
syncratic risk (idio) at the end of month t: In this Sorted Portfolios. In a next step, we examine
application, the forecast error is defined as the differ- whether statistically more accurate forecasts trans-
ence between expected and realized illiquidity over late into economic gains in a portfolio formation exer-
the next year within each decile portfolio. For the cise.17 In particular, we sort all bonds into decile
sake of brevity, we focus on comparisons of neural portfolios based on expected illiquidity at the end of
networks (nn_1) with the historical illiquidity–based each month t: Separately for each model and decile
approach (hist) and linear regressions (lm). Since the portfolio, we then calculate the equally weighted
nn_1 model produces slightly higher forecast errors mean of future realized illiquidity. Panel A of Table 6
(on average and over time) compared to both random presents the time-series averages of monthly portfo-
forests and gradient-boosted regression trees (see lio illiquidity (amihud measures). The last column adds
Table 5), this choice serves as a conservative lower results for the hypothetical case in which the sorting
bound for the following analysis. Figure 1 plots time- criterion is the bonds’ future realized illiquidity (real)
series averages of monthly forecast errors within all rather than a forecast model’s estimates, thus mim-
decile portfolios for the nn_1 model (grey bars) and icking perfect foresight. Panel B replicates the proce-
the benchmark models (red bars). We also include dure outlined above for each model but selects
the percentage differences in average forecast weights that differ from the equal weights to calcu-
errors relative to a benchmark model (black unfilled late the average illiquidity within each decile portfo-
squares), calculated as one minus the average MSE lio. In particular, the optimizer aims to minimize the
of the neural network divided by the average MSE of sum of squared deviations from the equal-weighting
the benchmark model. scheme, while requiring the portfolio-level rating (rat-
ing), yield (yield), and duration (dur) for the machine
For all forecast approaches, some of the extreme learning methods to be equal to those for the histori-
portfolios yield the largest average forecast errors. In cal benchmark model. This framework ensures a
particular, the expected illiquidity of bonds with (1) a straightforward comparison between machine learn-
high and low historical illiquidity, (2) a large and small ing–based methods and the historical illiquidity
amount outstanding, (3) a high rating, (4) a low expo- benchmark. It allows for more comparable decile
sure to bond market (systematic) risk, and (5) high portfolios and helps to avoid differences in expected
idiosyncratic risk are more difficult to predict. The illiquidity-sorted portfolios that are driven by differ-
graphs further suggest that neural networks reduce ences in their exposure to rating, yield, and duration.

114
 Predicting Corporate Bond Illiquidity via Machine Learning

Figure 1. Average Forecast Errors of Portfolio Sorts Based on Bond Characteristics

Notes: This figure plots the time-series averages of monthly mean squared forecast errors for decile portfolios based on bond char-
acteristics, that is, historical illiquidity (amihud_hist), size (size), rating (rating), systematic risk (beta), and idiosyncratic risk (idio) at the
end of each month t. The forecast error is defined as the difference between illiquidity forecasts and realized illiquidities over the
next year within each portfolio. Neural networks (nn_1) are compared with the historical illiquidity–based approach (hist) and linear
regressions (lm). The bars depict the time-series averages of monthly forecast errors within each decile portfolio for the nn_1 model
(grey bars) and the respective benchmark model (red bars). In addition, the percentage differences in average forecast errors relative
to the respective benchmark model are marked as black unfilled squares, calculated as one minus the average MSE of the neural
networks divided by the average MSE of the respective benchmark model. The baseline sample includes intraday transaction
records for the US corporate bond market reported in the enhanced version of Trade Reporting and Compliance Engine (TRACE)
for the sample period from July 2004 to November 2020.

Volume 80, Number 3 115

Financial Analysts Journal | A Publication of CFA Institute

Table 6. Illiquidity of Decile Portfolio Sorts

Forecast model
Reference
hist lm elanet rf gbrt nn_1 real

Panel A: Raw portfolio sorts

Low (L) 0.32 0.28 0.28 0.27 0.27 0.27 0.21
2 0.57 0.53 0.53 0.53 0.53 0.53 0.42
3 0.72 0.69 0.69 0.70 0.70 0.70 0.58
4 0.82 0.81 0.81 0.82 0.82 0.83 0.71
5 0.91 0.91 0.92 0.93 0.92 0.93 0.82
6 0.98 1.00 1.00 1.01 1.02 1.01 0.94
7 1.06 1.08 1.08 1.07 1.08 1.08 1.06
8 1.13 1.16 1.16 1.15 1.14 1.14 1.20
9 1.24 1.24 1.24 1.23 1.23 1.23 1.40
High (H) 1.44 1.46 1.47 1.47 1.47 1.46 1.83
H−L 1.12 1.18 1.19 1.20 1.20 1.19 1.62
t value − 7.01 12.16 11.37 10.11 5.83 −
Panel B: Portfolio sorts with controls for rating, yield, and duration
Low (L) 0.32 0.29 0.28 0.28 0.28 0.28 0.21
2 0.57 0.53 0.53 0.52 0.52 0.52 0.42
3 0.72 0.70 0.70 0.69 0.69 0.68 0.58
4 0.82 0.82 0.82 0.83 0.82 0.81 0.71
5 0.91 0.92 0.93 0.94 0.93 0.91 0.82
6 0.98 1.01 1.01 1.01 1.01 1.00 0.94
7 1.06 1.08 1.08 1.07 1.07 1.08 1.06
8 1.13 1.15 1.15 1.14 1.14 1.15 1.20
9 1.24 1.24 1.24 1.26 1.26 1.28 1.40
High (H) 1.44 1.45 1.47 1.48 1.50 1.53 1.83
H−L 1.12 1.17 1.19 1.20 1.22 1.24 1.62
t value − 3.69 7.29 6.72 10.18 10.63 −

Notes: This table examines differences in the predictive power of the different bond illiquidity forecast models (hist, lm, elanet, rf,
gbrt, and nn_1) from an economic perspective. Bonds are first sorted into decile portfolios based on illiquidity predictions (the
average monthly amihud measure) at the end of each month t: The equally weighted mean of future realized illiquidity is calculated
separately for each model and decile portfolio. Panel A presents the time-series averages of monthly figures. The last column adds
the corresponding results for the hypothetical case in which the sorting criterion is the bonds’ future realized illiquidity, that is,
assuming perfect foresight. Panel B replicates the procedure for each model but selects weights that slightly differ from the equal
weights to calculate the average illiquidity within each decile portfolio. The optimizer aims to minimize the sum of squared devia-
tions from the equal-weighting scheme, while requiring the portfolio-level rating (rating), yield (yield), and duration (dur) for the
machine learning methods to be equal to those for the hist model. H − L denotes the spread between the estimates in the high-
and low-illiquidity portfolios. A higher spread indicates that a given model is better at disentangling more liquid from less liquid
bonds, which suggests economic value added for investors in the form of transaction cost savings. The t values (using Newey–
West standard errors with 11 lags) are reported for the null hypothesis that the H − L illiquidity spread of a given column model
(lm, elanet, rf, gbrt, or nn_1) is not different from the H − L spread of the historical illiquidity benchmark (hist). The baseline sample
includes intraday transaction records for the US corporate bond market reported in the enhanced version of Trade Reporting and
Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

The results in Table 6 highlight that differences in resulting in positive average H–L spreads that are
statistical predictive performance translate into dif- statistically significant (not reported) and economi-
ferences in economic profitability. While the bench- cally large. Again, focusing on a comparison between
mark model and all machine learning models capture the historical illiquidity–based approach and the nn_1
the cross-sectional variation in realized illiquidity, model, we observe that for more liquid portfolios
their ability to disentangle more liquid from less liq- (e.g., decile 1), the average realized illiquidity for neu-
uid bonds differs. Average realized illiquidity within ral networks (0.27%) is lower and comes closer to
the decile portfolios obtained from expected illiquid- the true value of 0.21% than that for the hist model
ity line up monotonically with average realized illi- (0.32%). For less liquid portfolios (e.g., decile 10), the
quidity within the perfect-foresight decile portfolios, average realized illiquidity for the nn_1 model

116
 Predicting Corporate Bond Illiquidity via Machine Learning

(1.46%) is slightly higher and also closer to the true their exposure. Assuming perfect foresight, avoiding
value of 1.83% than that for the historical illiquidity– the 50% least liquid part of the market hypothetically
based approach (1.44%). This results in a 7 basis reduces the average price impact by a factor of 2.3
points larger H–L spread (1.19% for nn_1 vs. 1.12% (0.55% for deciles 1–5 vs. 1.29% for deciles 6–10
for hist). The differences between machine learning based on the averages in the column labelled
models and the benchmark model become more pro- “Reference” in Panel B of Table 6), improving portfo-
nounced after controlling for the decile portfolios’ lio turnover costs by 0:92% − 0:55% ¼ 0:37% per
exposures to rating, yield, and duration (especially for year. By allocating trading to the most liquid bonds
less liquid portfolios), resulting in an almost 11% (and avoiding the least liquid ones), for example, by
larger H–L spread (1.24% for nn_1 vs. 1.12% for hist). allocating a weight of 50% on the 10% most liquid
bonds, 25% on the second-most liquid, and so on,
Overall, machine learning techniques are better at the investor can even maintain her market impact
disentangling more liquid from less liquid bonds than costs below 50%  0:21% þ 25%  0:42% þ 12:5% 
the historical illiquidity benchmark, which suggests 0:58% þ 6:25%  0:71% þ 6:25%  0:82% ¼ 0:38%
economic value added to institutional investors.18 per year. Estimating the costs associated with specific
Due to the large transaction volumes in the corpo- securities is crucial for generating excess returns in
rate bond market, even the smallest improvements in classification strategies. Machine learning models are
illiquidity estimates will result in considerable transac- effective in this regard, surpassing the historical illi-
tion cost savings either directly in terms of a lower quidity–based prediction model and reducing
average bid–ask spread or indirectly in terms of a expected market impact by 12.5% (0.28% for the nn_
lower average price impact. Reduced transaction 1 model vs. 0.32% for the hist model) for the 10%
costs, in turn, have an immediate effect on improving and about 4% (0.64% for the nn_1 model vs. 0.67%
a portfolio’s risk–return profile. for the hist model using the allocation weights) for
the 50% most liquid bonds.
A Simple Example. To illustrate the importance
of illiquidity predictions for the performance of fixed- Finally, assume that the hypothetical investor wants to
income funds by way of an example that exploits the avoid the 50% least liquid bonds and concentrates
ranking performance of the different prediction mod- portfolio turnover on the most liquid bonds as outlined
els, take an average institutional investor with a bond above, but does not observe the real illiquidity distri-
portfolio size of $1 billion and a portfolio turnover of bution before trading. Relying on the historical illiquid-
5% per month. Moreover, assume that the average ity–based model would translate into 0.50% annual
bond portfolio consists of 2,000 bonds and that the cost (i.e., the average of deciles 1–5 in the column
portfolio’s annualized alpha is 1.0%. Ignoring other labeled “hist” in Table 6, Panel B). Therefore, the cost
transaction costs, without any price impact, this inves- of being unable to observe the future realized market
tor would be able to sell and buy bonds to rebalance impact ex ante is a 32% increase of the market impact
their portfolio for 5%  $1 billion ¼ $50 million of compared to the hypothetical perfect foresight sce-
bonds traded each month ($600 million of bonds nario of 0.38% (see above). The machine learning–
traded per year). Assuming a 0.92% Amihud (2002) based models can mitigate this cost by providing more
price impact (the average across deciles 1–10 in the accurate estimates of future illiquidity. For example,
right-most column labeled “Reference” in Panel B of the nn_1 model only leads to an increase of 21%
Table 6) results in a 0:92%  5% ¼ 4:6 bps reduction (0.46%) relative to the perfect-foresight case (0.38%),
in monthly alpha or 0.55% in annual alpha, wiping out that is, an 11 pps (32%−21%) reduction compared to
$5.5 million per year, that is, more than half of the the benchmark based on historical illiquidity (0.50%).
average annual gain of 1%  $1 billion ¼ $10 million
(before any other transaction costs). Cross-Sectional Bond Returns. A natural
extension of our analysis is to measure cross-sec-
The Amihud (2002) illiquidity measure is highly vari- tional bond returns. In particular, we again sort bonds
able across our bond universe, for example, the least into decile portfolios based on their historical illiquid-
liquid decile incurs a nearly nine times higher price ity or expected illiquidity at the end of each month t
impact than the most liquid one (1.83% vs. 0.21% in (using the hist, lm, and nn_1 models). We then com-
Table 6). Better illiquidity predictions help to control pute the portfolio return in the next month t þ 1. The
average turnover costs by enabling investors to focus results for the full sample and three subperiods are
on the most liquid decile portfolios (and sorting out shown in Table 7. As expected, bonds in decile 10
the least liquid decile portfolios) when rebalancing (more illiquid bonds) outperform bonds in decile 1

Volume 80, Number 3 117

Financial Analysts Journal | A Publication of CFA Institute

Table 7. Returns of Decile Portfolio Sorts

2012–2020 2012–2014 2015–2017 2018–2020
hist lm nn_1 hist lm nn_1 hist lm nn_1 hist lm nn_1

Low (L) 0.38 0.37 0.36 0.55 0.53 0.48 0.27 0.29 0.30 0.34 0.32 0.31
2 0.38 0.38 0.39 0.50 0.54 0.55 0.29 0.30 0.31 0.36 0.34 0.32
3 0.40 0.38 0.40 0.53 0.50 0.52 0.32 0.31 0.35 0.36 0.33 0.36
4 0.41 0.40 0.40 0.51 0.47 0.48 0.36 0.38 0.37 0.38 0.35 0.36
5 0.45 0.40 0.42 0.53 0.50 0.52 0.47 0.37 0.39 0.37 0.35 0.35
6 0.46 0.44 0.42 0.51 0.54 0.51 0.48 0.44 0.39 0.40 0.35 0.39
7 0.43 0.46 0.47 0.50 0.50 0.52 0.42 0.48 0.46 0.39 0.41 0.44
8 0.49 0.48 0.47 0.52 0.51 0.47 0.52 0.50 0.48 0.42 0.43 0.44
9 0.49 0.53 0.51 0.54 0.57 0.56 0.53 0.54 0.53 0.41 0.48 0.44
High (H) 0.58 0.65 0.64 0.65 0.71 0.72 0.64 0.70 0.73 0.46 0.53 0.49
H−L 0.20 0.27 0.29 0.10 0.19 0.25 0.37 0.41 0.43 0.12 0.21 0.18

Notes: This table presents differences in the predictive power of the different bond illiquidity forecast models (hist, lm, and nn_1) for
cross-sectional bond returns. Bonds are first sorted into decile portfolios based on their historical illiquidity or expected illiquidity
(using the average monthly amihud measure) at the end of each month t: In a second step, the equally weighted return is calculated
for each prediction model and decile portfolio in the next month t þ 1 (in % per month). H − L denotes the spread between the esti-
mates in the high- and low-illiquidity portfolios. A higher spread indicates that a given model generates a higher illiquidity premium
in the cross-section of bond returns. Returns are reported for the full test sample and three subperiods (2011–2013, 2014–2016,
2017–2019). The baseline sample includes intraday transaction records for the US corporate bond market reported in the enhanced
version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

(less illiquid bonds). Even more important from an nonlinear machine learning models with each other.
asset pricing perspective, the H − L spread is higher For example, in Panel A of Table 6, the difference in
for the machine learning–based prediction models (lm the H–L spreads between the lm (1.18) and nn_1
and nn_1) compared to the historical illiquidity model (1.19) methods is negligible, but it becomes larger
(hist). The same results continue to hold for bond when appropriately controlling for risk in Panel B
returns over the next 12 months (not reported). (1.17 for lm vs. 1.24 for nn_1). In Table 7, the lm
Prediction models using machine learning techniques model even generates a slightly higher illiquidity pre-
generate a higher illiquidity premium in the cross- mium than the nn_1 model during the last subperiod
section of bond returns than the benchmark model. (2018–2020). In all other subperiods, the nn_1 model
These results support Amihud and Mendelson’s dominates the lm model marginally. These patterns
(1986) insight that investors require higher returns are important for the practical implementation in
for more illiquid bonds to compensate them for their portfolio management because neural networks in
higher trading expenses. particular are computationally extremely costly.

Practical Implications. Comparing our findings In light of these findings, whether the complexity and
for the statistical assessment with the economic per- resourcefulness of more sophisticated machine learn-
formance of our different forecast methods reveals ing methods is justified in the asset management
another issue that seems particularly important from practice most likely depends on the specific applica-
a practitioner’s perspective. While the results in tion. If illiquidity predictions are merely used to rank
Tables 4 and 5 suggest that nonlinear machine learn- bonds and sort out the least liquid ones, as illustrated
ing models (rf, gbrt, and nn_1) strongly outperform in the example above, models that incorporate a set
both the historical benchmark (hist) and their linear of predictor variables (in addition to historical illiquid-
counterparts (lm and elanet) when it comes to level ity) in a linear way seem satisfactory and are straight-
predictions of illiquidity, as measured by a statistical forward to implement. However, in contrast to such
comparison of their forecast errors, the results are relatively simple ranking and/or sorting exercises,
more nuanced for the mere rank forecasts. In gen- there are many use cases that require the highest
eral, machine learning methods perform better than possible accuracy of illiquidity level predictions. In
the historical illiquidity benchmark in sorting bonds particular, practical applications that involve bond
into expected illiquidity portfolios, but the difference portfolio optimization under the constraint to mini-
becomes less pronounced when comparing linear and mize transaction costs should benefit from more

118
 Predicting Corporate Bond Illiquidity via Machine Learning

complex machine learning methods that account for the contributions of individual variables using relative
nonlinearity and multiway interactions. Furthermore, variable importance metrics and explore patterns of
the benefits from nonlinear models may be more nonlinear and interactive effects in the relationship
important at some times than at others. For example, between predictor variables and illiquidity estimates.
Drobetz et al. (2024) show that more complex mod-
els are required during turbulent times when predic- Variable Importance. We begin with investigat-
tions become more difficult. ing which variables are, on average, most important
for the predictions obtained from linear regressions
To provide a specific example, we note that accurate (lm) and neural networks (nn_1). Given that we re-
forecasts of corporate bond illiquidity are highly estimate our models on an annual basis, it is also
important from a regulatory perspective. The “SEC instructive to inspect whether a predictor’s contribu-
Liquidity Rule” requires that 85% of a fund could be tion to the overall forecast ability of a model changes
liquidated in fewer than five days with a maximum over time. Separately for each model and re-estimation
participation of 20% of daily dollar trading volumes
date, we compute the variable importance matrix using
to be applied to a corporate bond portfolio. If the
a two-step approach: First, we compute the absolute
fund mimics the performance of a corporate bond
variable importance as the increase in MSE from
index, as most exchange-traded funds attempt to do,
randomly permuting the values of a given predictor
the portfolio construction process should be viewed
variable in the training sample. Second, we normalize
as a tracking error minimization under some liquidity
the absolute variable importance measures to sum to
and capacity constraints. While capacity constraints
one, signaling the relative contribution of each variable
may be based on estimates of the bonds’ trading vol-
to the lm and nn_1 model.
umes, controlling for turnover costs depends more
on price impact measures, such as the Amihud (2002) Analyzing the relative variable importance metrics
illiquidity measure. Corporate bond ETFs are known over time, more volatile metrics indicate that all
to achieve lower Sharpe ratios because such instru- covariates in the predictor set should be considered
ments must pay for liquidity (Houweling 2011). As a important. In contrast, stable metrics mean we should
result, the ability to accurately forecast bond illiquid- remove uninformative predictors permanently, as
ity is crucial for improving capacity and turnover they may decrease a model’s signal-to-noise ratio.
costs of such replicating strategies and, based on our Figure 2 depicts the relative variable importance met-
statistical analysis of level forecast errors, more com- rics over the sample period for linear regressions
plex machine learning–based models seem to be (Panel A) and neural networks (Panel B). To allow for
most appropriate to accomplish this task. More gen- better visual assessment, we omit the bars for the
erally, this argument is true for any mutual bond historical illiquidity predictor. The relative variable
fund that has achieved a certain size. Because regula-
importance of amihud_hist can be inferred by sub-
tory liquidity requirements must be met at all times,
tracting the aggregate relative importance of all other
this can prove to be difficult as funds become large,
predictors from one. On average, both models place
unless they are willing to pay or make their investors
the largest weight on historical illiquidity; this predic-
pay for liquidity.
tor accounts for more than 40% of the aggregate
average variable importance for the lm model but
only around 35% for the nn_1 model. On the one
Characteristics and Functioning
hand, high weights are expected because realized illi-
Scheme of Machine Learning quidity is persistent and has long-memory properties.
Estimators On the other hand, the lower weight placed on his-
Recognizing that machine learning–based models torical illiquidity by neural networks relative to linear
outperform the historical illiquidity benchmark and regressions helps to explain why the nn_1 model out-
that nonlinearity as well as interactions can further performs the lm model in terms of lower forecast
help in accurately modeling expected bond illiquidity, errors in general, but especially within extreme
we now focus on determining how these techniques, decile portfolios sorted on historical illiquidity (see
which are often referred to as “black boxes,” achieve Figure 1). Linear regressions miss out on extracting
outperformance. This black box problem is addressed valuable information from the nonlinear and interac-
by examining the characteristics and functioning tive patterns in the relationship between our set of
scheme of neural networks, focusing particularly on fundamental as well as macroeconomic predictor
the nn_1 model.19 We decompose predictions into variables and expected illiquidity.

Volume 80, Number 3 119

Financial Analysts Journal | A Publication of CFA Institute

Figure 2. Variable Importance

Notes: This figure shows the relative importance of the variables included as predictors in linear regressions (Panel A) and neural net-
works (Panel B) at each re-estimation date. For this purpose, the relative variable importance matrix is calculated based on a two-step
approach: First, the absolute variable importance is computed as the reduction in R2 from setting all values of a given predictor to zero
within the training sample. Second, the absolute variable importance measures are normalized to sum to 1, signaling the relative contribu-
tion of each variable to a model. The baseline sample includes intraday transaction records for the US corporate bond market reported
in the enhanced version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

120
 Predicting Corporate Bond Illiquidity via Machine Learning

Among the remaining variables, the two models iden- Panel A of Figure 3 illustrates the marginal associa-
tify slightly different predictors as most relevant for tion between dur and IFi, tþkjt for linear regressions
estimating illiquidity. While the lm model considers (dashed line) and neural networks (solid line), respec-
the default spread (dfy) highly informative, neural tively. We add a histogram that depicts the historical
networks predominantly extract information from distribution of dur. This visualization allows us to
fundamental predictors. In particular, the bond-level assess the empirical relevance of differences in pre-
predictor duration (dur), size (size), and maturity (mat) dictions obtained from the lm and nn_1 models for
are most important for the nn_1 model, accounting the overall forecast results. At the left end of the dis-
for roughly 18%, 18%, and 14% of the aggregate tribution, approximately within the (þ1:8, þ 6:3)
average variable importance, respectively. In contrast, interval, the predictions obtained from linear regres-
the default spread (dfy) is much less informative for sions and neural networks are similar. We identify an
neural network models. Accordingly, given the supe- increasing linear relationship between dur and IFi, tþkjt
rior predictive performance of the nn_1 model, the for the lm model and a close-to-linear relationship for
time variation in bond illiquidity is driven more by the nn_1 model, suggesting that bonds with a shorter
changes in bond characteristics than by changes in duration are more liquid than their medium-duration
the underlying economic conditions. Moreover, while counterparts. However, outside this interval, the mar-
most variables have some relevance based on their ginal association between duration and expected illi-
average metrics, the analysis reveals that these met- quidity delineated in the neural network model is
rics change notably over time. Because this time vari- strongly negative, suggesting that bonds with a lon-
ation is apparent for all predictor variables, we ger duration are also likely to be more liquid than
conclude that each variable is an important contribu- their medium-duration counterparts. Overall, this
tor in all models, albeit to varying degrees. Overall, leads to a nonsymmetrical inverted U-shaped rela-
the variable importance results in Figure 2 do not tionship for the nn_1 model. In sharp contrast, the lm
recommend that we should remove specific model, by construction, must continue to follow the
predictors.20 increasing linear relationship over the entire range of
duration, resulting in illiquidity estimates from linear
Nonlinearity and Interactions. Tree-based regressions that diverge substantially from those
models and neural networks are superior to the his- obtained using neural networks.
torical illiquidity benchmark, and they also tend to
The nonsymmetrical inverted U-shaped relationship
outperform linear regressions with the same set of
between dur and IFi, tþkjt (compared to an increasing
covariates. A large part of this outperformance must
be attributable to their ability to exploit nonlinear linear relationship) is more consistent with (1) the
empirical patterns observed when plotting duration
and interactive patterns in the relationship between
and realized illiquidity during the out-of-sample
predictors and expected bond illiquidity. Therefore, in
period simultaneously (not reported) and (2) anec-
a final step, we analyze in more detail whether and
dotal evidence. Anecdotal evidence is twofold. First,
how neural networks (nn_1) capture nonlinearity and
the number of institutional investors with a natural
interactions. For comparison, we contrast the results
duration target (e.g., property and casualty vs. life
with illiquidity estimates obtained from linear regres-
insurance companies) is higher for both ends of the
sions (lm).21
yield curve because they attempt to match their
We first examine the marginal association between a short-term and long-term liabilities, respectively.
single predictor variable and its illiquidity predictions Therefore, these types of bonds, that is, with either
(IFi, tþkjt , with k ¼ 12). As an example, we select a a lower or a higher duration, are issued more fre-
quently (consistent with the historical distribution of
bond’s duration (dur), one of the most influential pre-
dur depicted by the histogram). Importantly, they are
dictors in our analysis (see Figure 2). To visualize the
also traded more frequently, thereby increasing their
average effect of dur on IFi, tþkjt , we set all predictors liquidity. Second, institutional investors are likely to
to their uninformative median values within the train- increase (decrease) their portfolio’s duration by buy-
ing sample at each re-estimation date. We then vary ing high-duration (low-duration) bonds when they
dur across the minimum and maximum values of its expect interest rates to rise (fall). This behavior is
historical distribution and compute the expected illi- another reason for lower- and higher-duration bonds
quidity. Finally, we average the illiquidity predictions to be traded more frequently, increasing their liquid-
across all re-estimation dates. ity further. Because a considerable share of our

Volume 80, Number 3 121

Financial Analysts Journal | A Publication of CFA Institute

Figure 3. Nonlinear and Interactive Effects in Estimating Corporate Bond Illiquidity

Notes: This figure examines the models’ ability to capture nonlinear and interactive effects in estimating future bond illiquidity (the average
monthly amihud measure). Panel A illustrates the marginal association between bond duration (dur) and its illiquidity estimates (IFi, tþkjt , with
k ¼ 12) for linear regressions (dashed line) and neural networks (solid line), respectively. It also shows a histogram that depicts the histori-
cal distribution of dur. To visualize the average effect of dur on IFi, tþkjt , all predictors are set to their uninformative median values within
the training sample at each re-estimation date. In the next step, dur is varied over the minimum and maximum values of its historical distri-
bution, and the illiquidity estimates are computed. Finally, the average illiquidity estimates across all re-estimation dates are estimated and
presented in the panel. Panel B shows the interactive effect between dur and bond rating (rating) on IFi, tþkjt : To this end, the procedure
outlined above is replicated, but this time the illiquidity estimates are computed for different levels of rating across its minimum and maxi-
mum values. Low and high levels for rating are marked with red and green lines; dashed and solid lines refer to linear regressions (lm) and
neural networks (nn_1), respectively. The baseline sample includes intraday transaction records for the US corporate bond market reported
in the enhanced version of Trade Reporting and Compliance Engine (TRACE) for the sample period from July 2004 to November 2020.

122
 Predicting Corporate Bond Illiquidity via Machine Learning

observations lies within the lower and upper parts of predominantly during a sample period with historically
the historical distribution, the differences in predic- low interest rates. In this “zero lower bound” environ-
tions are practically relevant. Our analysis highlights ment, many institutional investors adapted to yield
the need to allow for nonlinear impacts of the pre- scarcity by taking on more risk, that is, they shifted
dictor variables on expected illiquidity. We further their focus to lower-rated bonds to meet their need for
note that nonlinear relationships (both U- and S- income. With respect to duration, they often chose
shaped) are similarly observable for other predictors shorter-duration bonds with lower ratings, for example,
(not reported), for example, a bond’s historical illi- high-yield bonds, as opposed to their higher-rating
quidity (amihud_hist), maturity (mat), size (size), and counterparts. This change in preferences has led to a
age (age). relative shift in demand, which may have contributed
to a decrease in the liquidity of bonds with a higher
Next, we investigate between-predictor interactions rating.
in estimating corporate bond illiquidity, referring
again to dur as our baseline covariate. In addition, we Taken together, these visualizations provide an expla-
select rating, another highly influential predictor (see nation for our main finding that more complex
Figure 2), as our interactive counterpart and replicate machine learning models, such as regression trees
the procedure just described. In this case, we com- and neural networks, are able to generate more accu-
pute expected illiquidity for different levels of rating rate bond illiquidity forecasts than their linear coun-
across its minimum and maximum values. The inter- terparts. Linear regressions (both simple and
active effect between dur and rating on IFi, tþkjt is illus- penalized), by construction, cannot capture nonlinear
trated in Panel B of Figure 3. Low and high levels for and multiway interactive effects that seem to
rating are marked as red and green lines, respectively. describe real-world phenomena in a much better
If there is no interaction, or if the model is unable to way. In this light, our analysis helps to explain the
capture interactions, computing expected illiquidity outperformance of regression trees and neural net-
for different levels of rating shifts the lines from works (which “learn” these complex patterns from
Panel A up- or downward in parallel. In this case, the the training and validation data) over the historical
distance between the lines is identical for any given illiquidity benchmark and linear regression models in
value of dur. This pattern is apparent for linear terms of lower forecast errors.
regressions (drawn as dotted lines), because no pre-
specified interaction term, for example, dur  rating
for the interaction between dur and rating, is included Conclusion
as a predictor in the linear regression framework. Understanding a bond’s multi-faceted liquidity char-
The dotted lines are shifted downward when rating acteristics and predicting bond illiquidity are relevant
increases, indicating that an increase in rating that is topics from an asset pricing point of view, but they
independent of the bond’s duration decreases IFi, tþkjt : are equally important from a regulatory and real-
word investor perspective. Our paper contributes to
For neural networks, the same pattern is only observ- a better understanding of the characteristics of cor-
able for the right end of the dur distribution. In con- porate bond illiquidity and of how to transform this
trast, at the left end of the distribution, unlike the lm information into reliable illiquidity forecasts. In partic-
model, the nn_1 model uncovers interactive effects ular, we compare the predictive performance of
between dur and a bond’s rating in predicting illiquid- machine learning–based illiquidity estimators (linear
ity.22 This interactive effect is so strong that it reverses regressions, tree-based models, and neural networks)
the isolated effects of duration and rating on expected to that of the historical illiquidity benchmark, which
illiquidity, that is, bonds with a shorter duration and, at is the most commonly used model. All machine learn-
the same time, a higher rating tend to be less liquid ing models outperform the historical illiquidity–based
than their lower-rating counterparts. This finding is approach from both a statistical and an economic
again consistent with anecdotal evidence. Liquidity of perspective. The outperformance is attributable to
high-yield bonds is concentrated at the short-term part these models’ ability to exploit information from a
of the curve because these bonds tend to have shorter large set of bond characteristics that impact bond illi-
durations. On average, however, bond liquidity quidity. Tree-based models (random forests and gra-
decreases with higher credit risk. The reverse is true dient-boosted regression trees) and neural networks
for the shorter-term part of the curve, which one perform similarly and work remarkably well. These
could perceive as counterintuitive. An explanation is more complex approaches outperform linear regres-
that our machine learning models have been fitted sions with the same set of covariates, particularly in

Volume 80, Number 3 123

Financial Analysts Journal | A Publication of CFA Institute

terms of prediction level accuracy, because of their illiquidity as opposed to bond portfolio optimizations
ability to utilize nonlinear and interactive patterns. that require level forecasts of illiquidity. An obvious
From a practitioner’s perspective, our results suggest open question is whether our findings can be trans-
that the choice of the appropriate machine learning ferred to corporate bonds for which historical illiquid-
model depends on the specific application, such as ity data are not readily available. We leave this task
simple bond rankings and sortings based on expected for future research.

Editor's Note
Submitted 24 June 2023
Accepted 22 April 2024 by William N. Goetzmann

Notes
1. Other papers in this research area that find evidence 8. Because we apply on a one-year forecast horizon, there is
for superior stock selection based on a large set of a one-year gap between the end of the sample that is
predictors are Rasekhschaffe and Jones (2019), Chen, used for training and validation (2010:06) and the
Pelger, and Zhou (2022), and Bryzgalova, Pelger, and estimation date (2011:06).
Zhou (2023). Related studies document similar results for
international data (Tobek and Hronec 2021), European 9. Neural network models are computationally intensive and
data (Drobetz and Otto 2021), emerging markets data can be specified in innumerable different architectures. We
retreat from tuning parameters (e.g., the size of batches or
(Hanauer and Kalsbach 2023), Chinese data (Leippold
the number of epochs) and specify five different models,
et al., 2023), and crash prediction models (Dichtl,
assuming that our nn_1-nn_5 architectures are a conservative
Drobetz, and Otto 2023).
lower bound for the predictive performance of neural
2. Reichenbacher, Schuster, and Uhrig-Homburg (2020) also network models. Because the predictive performance of
use a large set of predictors for expected bond liquidity, neural network models deteriorates slightly in the number of
but they work with the linearity assumption in their hidden layers in our application (not reported), we only
estimation model. present the results for the nn_1 architecture.

3. We use the enhanced version of TRACE instead of the 10. Following Becker et al. (2021), we omit bonds with fewer
standard version because it additionally contains than 50 illiquidity estimates to allow for valid inference.
uncapped transaction volumes and information on 11. The cross-sectional and time-series means for realized
whether the trade is a buy, a sell, or an interdealer illiquidity are −4.75 and −4.78, respectively (not reported).
transaction. This refinement enables us to construct
measures that capture different aspects of bond illiquidity 12. In most economic applications, when comparing different
based on intraday bond transactions. models, a single model does not exist that significantly
dominates all competitors because the data are not
4. The detailed transaction data allow us to compute direct sufficiently informative to provide an unequivocal answer.
liquidity measures as opposed to indirect measures based However, it is possible to reduce the set of models to a
on bond characteristics and/or end-of-day prices smaller set of models—the so-called model confidence set
(Houweling, Mentink, and Vorst 2005). (MCS)—that contains the best model(s) with a given level
of confidence. Hansen, Lunde, and Nason’s (2011) MCS
5. To control for return outliers not driven by illiquidity, we determines the set of models that composes the best
omit observations with daily amihud measures exceeding 5% model(s) from a collection of models, where “best” is
on a given day. Our main results remain qualitatively similar defined in terms of the MSE. Informative data will result
when using other cut-off thresholds, e.g., 1% or 10%. in a MCS that contains only the best model. Less
informative data make it difficult to distinguish between
6. Because no prior study has examined whether this set of
models and result in a MCS that contains several models.
variables is helpful for predicting bond illiquidity, a potential
In our applications, we examine statistical significance at
lookahead bias should not be an issue in our analysis.
the 5% level, translating into 95% model confidence sets.
7. Alternatively, one-month and five-year forecast horizons
13. According to Gu, Kelly, and Xiu (2020), DM test statistics are
are common in the literature (k ¼ 1 and k ¼ 60,
asymptotically Nð0, 1Þ-distributed and test the null hypothesis
respectively). Both alternatives have shortcomings in our that the divergence between two models is zero. They map
setup, which is why we opt for a one-year forecast to p-values in the same way as regression t-statistics.
horizon. First, one-month illiquidity measures are very
noisy, which hampers the evaluation of forecast errors. 14. Due to limited data availability, the test sample in our
Second, forecast horizons much longer than 12 months baseline setting does not contain the 2007–2008 global
are less common in the industry due to the underlying financial crisis, during which the availability of credit
nature of fiscal years. suddenly plummeted. In a robustness test (not reported),

124
 Predicting Corporate Bond Illiquidity via Machine Learning

we shorten the length of the sample used for training and measures. Machine learning–based methods produce a
validation to three years in order to include the global superior misclassification distribution, which may translate
financial crisis in the test sample. Again, all machine into economic outperformance.
learning models dominate the historical illiquidity–based
model during this severe crisis. 18. Table 6 also contains the t-statistics (using Newey-West
standard errors with 11 lags) for the null hypothesis that
15. Table C2 in Online Supplemental Appendix C confirms the H–L illiquidity spread of a given column model (lm,
that this conclusion remains robust for our two other elanet, rf, gbrt, or nn_1) is not different from the H–L
illiquidity measures: t_volume and t_spread. To check spread of the historical illiquidity-based model (hist). All t-
robustness even further, this table also contains the results statistics for pairwise differences in portfolio means
for two additional illiquidity measures: First, we apply indicate statistical significance, i.e., the null hypothesis of
Lesmond, Ogden, and Trzcinka’s (1999) illiquidity measure indifference can be rejected in all cases, thus confirming
based on zero daily bond returns (p_zeros), where a larger that machine learning techniques are reliably better at
fraction of zero returns in a given sample month indicates disentangling more liquid from less liquid bonds.
lower liquidity. Second, we use Roll’s (1984) implicit measure
of the bid–ask spread based on the covariance of daily bond 19. Similar to random forests and gradient-boosted regression
trees, neural networks exhibit low forecast errors (both
returns and their lagged returns (Roll’s spread). To this end,
on average and over time; see Table 5), produce accurate
we calculate the negative autocorrelation of bond returns
forecasts for bonds with extreme characteristics (see
within a given sample month, with higher numbers indicating
Figure 1), and perform well in a portfolio formation
lower liquidity. The results are qualitatively similar, albeit the
exercise (see Table 6).
performance advantage of machine learning models is less
pronounced for the p_zeros measure. 20. To be on the conservative side, we compare the statistical
and economic predictive performance of the original nn_1
16. In a robustness test, we implement a Giacomini–White
model with versions that only consider the top 5 or 10
(2006) test for conditional predictive performance. The DM
predictors in terms of their relative variable importance. Out-
test is unconditional in the sense that it asks which forecast
of-sample test results (not reported) suggest that no model
was more accurate, on average, in the past; it may thus be
version exhibits superior outperformance in any of these
appropriate for making recommendations about which
tests, so we choose not to remove unconditionally less
forecast may be better for an unspecified future date. As
informative variables from the predictor set and instead
elaborated in Giacomini and White (2006), “the conditional
consider each predictor as informative (albeit to varying
approach asks instead whether we can use available
degrees). Furthermore, we caution that the pre-estimation
information—above and beyond past average behavior—to
variable selection based on relative importance metrics
predict which forecast will be more accurate for a specific derived from the entire sample period could lead to foresight
future date” (p. 1547). To describe the specific future date, bias, undermining the credibility of out-of-sample tests.
we use the default spread at the end of the last month as
the conditioning variable that captures the prevailing state of 21. The patterns and their implications are qualitatively similar
the economy. The results are presented in Table C3 in when comparing gradient-boosted regression trees (gbrt) and
Online Supplemental Appendix C. All test statistics indicate neural networks (nn_1) to estimates obtained from penalized
statistical significance. Overall, the results support the DM linear regressions (elanet). This result confirms that the ability
tests. Machine learning–based forecast methods outperform to exploit nonlinear and interactive patterns leads to the
the historical illiquidity benchmark not only in terms of their outperformance of tree-based models and neural networks
unconditional predictive ability but also in terms of their over linear regressions.
conditional predictive ability.
22. Despite being slightly less pronounced, the nn_1 model
17. Online Supplemental Appendix B presents an alternative also reveals interactive effects between other bond
test for classification performance based on confusion characteristics in estimating future illiquidity, for example,
matrices that contrasts predicted and realized classes, between a bond’s historical illiquidity (amihud_hist) and
together with accuracy and ranking loss as classification size (size).

References
Amihud, Y. 2002. “Illiquidity and Stock Returns: Cross-Section Bali, T. G., A. Goyal, D. Huang, F. Jiang, and Q. Wen. 2020.
and Time-Series Effects.” Journal of Financial Markets 5 (1): 31– Predicting Corporate Bond Returns: Merton Meets Machine
56. doi:10.1016/S1386-4181(01)00024-6. Learning. Georgetown McDonough School of Business
Research Paper No. 3686164.
Amihud, Y., and H. Mendelson. 1986. “Asset Pricing and the
Bid-Ask Spread.” Journal of Financial Economics 17 (2): 223– Bali, T. G., A. Goyal, D. Huang, F. Jiang, and Q. Wen. 2022.
249. doi:10.1016/0304-405X(86)90065-6. The Cross-Sectional Pricing of Corporate Bonds Using Big Data
and Machine Learning, Working Paper.
Avramov, D., S. Cheng, and L. Metzker. 2022. “Machine
Learning versus Economic Restrictions: Evidence from Stock Bali, T. G., A. Subrahmanyam, and Q. Wen. 2021. “Long-Term
Return Predictability.” Management Science 69 (5): 2587–2619. Reversals in the Corporate Bond Market.” Journal of Financial
doi:10.1287/mnsc.2022.4449. Economics 139 (2): 656–677. doi:10.1016/j.jfineco.2020.08.007.

Volume 80, Number 3 125

Financial Analysts Journal | A Publication of CFA Institute

Bao, J., J. Pan, and J. Wang. 2011. “The Illiquidity of Corporate Crisis.” Journal of Financial Economics 103 (3): 471–492. doi:10.
Bonds.” The Journal of Finance 66 (3): 911–946. doi:10.1111/j. 1016/j.jfineco.2011.10.009.
1540-6261.2011.01655.x.
Diebold, F., and R. Mariano. 1995. “Comparing Predictive
Becker, J., F. Hollstein, M. Prokopczuk, and P. Sibbertsen. Accuracy.” Journal of Business & Economic Statistics 13 (3): 253–
2021. “The Memory of Beta.” Journal of Banking & Finance 124 263. doi:10.1080/07350015.1995.10524599.
(1): 106026. doi:10.1016/j.jbankfin.2020.106026.
Drobetz, W., F. Hollstein, T. Otto, and M. Prokopczuk. 2024.
Bessembinder, H., K. M. Kahle, W. F. Maxwell, and D. Xu. “Estimating Stock Market Betas via Machine Learning.” Journal
2009. “Measuring Abnormal Bond Performance.” Review of of Financial and Quantitative Analysis 1–56. doi:10.1017/
Financial Studies 22 (10): 4219–4258. doi:10.1093/rfs/hhn105. S0022109024000036.
Bessembinder, H., W. Maxwell, and K. Venkataraman. 2006. Drobetz, W., and T. Otto. 2021. “Empirical Asset Pricing via
“Market Transparency, Liquidity Externalities, and Institutional Machine Learning: Evidece from the European Stock Market.”
Trading Costs in Corporate Bonds.” Journal of Financial Journal of Asset Management 22 (7): 507–538. doi:10.1057/
Economics 82 (2): 251–288. doi:10.1016/j.jfineco.2005.10.002. s41260-021-00237-x.
Bianchi, D., M. Bu€chner, and A. Tamoni. 2021. “Bond Risk Fedenia, M., S. Nam, and T. Ronen. 2021. Machine Learning in
Premiums with Machine Learning.” The Review of Financial the Corporate Bond Market and Beyond: A New Classifier,
Studies 34 (2): 1046–1089. doi:10.1093/rfs/hhaa062. SSRN Working Paper.
Blitz, D., T. Hoogteijling, H. Lohre, and P. Messow. 2023. Freyberger, J., A. Neuhierl, and M. Weber. 2020. “Dissecting
“How Can Machine Learning Advance Quantitative Asset Characteristics Nonparametrically.” The Review of Financial
Management.” Journal of Portfolio Management 49 (9): 78–95. Studies 33 (5): 2326–2377. doi:10.1093/rfs/hhz123.
Bongaerts, D., F. de Jong, and J. Driessen. 2017. “An Asset Friewald, N., R. Jankowitsch, and M. Subrahmanyam. 2012.
Pricing Approach to Liquidity Effects in Corporate Bond
“Illiquidity or Credit Deterioration: A Study of Liquidity in the
Markets.” The Review of Financial Studies 30 (4): 1229–1269.
US Corporate Bond Market during Financial Crises.” Journal of
doi:10.1093/rfs/hhx005.
Financial Economics 105 (1): 18–36. doi:10.1016/j.jfineco.2012.
Bryzgalova, S., M. Pelger, and J. Zhu. 2023. “Forest through the 02.001.
Trees: Building Cross-Sections of Stock Returns.” Journal of
Giacomini, R., and H. White. 2006. “Tests of Conditional
Finance. Forthcoming.
Predictive Ability.” Econometrica 74 (6): 1545–1578. doi:10.
Chakravarty, S., and A. Sarkar. 2003. “Trading Costs in Three 1111/j.1468-0262.2006.00718.x.
U.S. Bond Markets.” The Journal of Fixed Income 13 (1): 39–48.
Gu, S., B. Kelly, and D. Xiu. 2020. “Empirical Asset Pricing via
doi:10.3905/jfi.2003.319345.
Machine Learning.” The Review of Financial Studies 33 (5):
Chen, L., M. Pelger, and J. Zhou. 2022. “Deep Learning in Asset 2223–2273. doi:10.1093/rfs/hhaa009.
Pricing.” Management Science. Forthcoming.
Hanauer, M., and T. Kalsbach. 2023. “Machine Learning and
Cherief, A., M. Ben Slimane, J.-M. Dumas, and H. Fredj. 2022. the Cross-Section of Emerging Market Stock Returns.” Emerging
“Credit Factor Investing with Machine Learning Techniques.” Markets Review 55: 101022. doi:10.1016/j.ememar.2023.
Working paper 128-2022, Amundi Asset Management. 101022.
Chordia, T., A. Sarkar, and A. Subrahmanyam. 2005. “An Hansen, P. R., A. Lunde, and J. M. Nason. 2011. “The Model
Empirical Analysis of Stock and Bond Market Liquidity.” The Confidence Set.” Econometrica 79 (2): 453–97.
Review of Financial Studies 18 (1): 85–129. http://www.jstor.
org/stable/3598068. Hong, G., and A. Warga. 2000. “An Empirical Study of Bond
Market Transactions.” Financial Analysts Journal 56 (2): 32–46.
Cosemans, M., R. Frehen, P. C. Schotman, and R. Bauer. 2016. doi:10.2469/faj.v56.n2.2342.
“Estimating Market Betas Using Prior Information Based on
Firm Fundamentals.” Review of Financial Studies 29 (4): 1072– Hotchkiss, E., and G. Jostova. 2017. “Determinants of
1112. doi:10.1093/rfs/hhv131. Corporate Bond Trading: A Comprehensive Analysis.” Quarterly
Journal of Finance 07 (02): 1750003. doi:10.1142/
Dichtl, H., W. Drobetz, and T. Otto. 2023. “Forecasting Stock S2010139217500033.
Market Crashes via Machine Learning.” Journal of Financial
Stability 65: 101099. doi:10.1016/j.jfs.2022.101099. Houweling, P. 2011. “On the Performance of Fixed Income
Exchange Traded Funds.” The Journal of Index Investing 3 (1):
Dickerson, A., P. Mueller, and C. Robotti. 2023. “Priced Risk in 39–44. doi:10.3905/jii.2012.3.1.039.
Corporate Bonds.” Journal of Financial Economics 150 (2):
103707. doi:10.1016/j.jfineco.2023.103707. Houweling, P., A. Mentink, and T. Vorst. 2005. “Comparing
Possible Proxies of Corporate Bond Liquidity.” Journal of
Dick-Nielsen, J. 2009. “Liquidity Biases in TRACE.” The Journal
Banking & Finance 29 (6): 1331–1358. doi:10.1016/j.jbankfin.
of Fixed Income 19 (2): 43–55. doi:10.3905/jfi.2009.19.2.043.
2004.04.007.
Dick-Nielsen, J. 2014. How to Clean Enhanced TRACE Data,
Jankowitsch, R., A. Nashikkar, and M. Subrahmanyam. 2011.
Working Paper.
“Priced Dispersion in OTC Markets: A New Measure of
€tter, and D. Lando. 2012. “Corporate
Dick-Nielsen, J., P. Feldhu Liquidity.” Journal of Banking & Finance 35 (2): 343–357. doi:10.
Bond Liquidity before and after the on-Set of the Subprime 1016/j.jbankfin.2010.08.016.

126
 Predicting Corporate Bond Illiquidity via Machine Learning

Kaufmann, H., P. Messow, and J. Vogt. 2021. “Boosting the Liquidity, with an Application to Corporate Bonds.” Journal of
Equity Momentum Factor in Credit.” Financial Analysts Journal Financial Economics 88 (2): 272–298. doi:10.1016/j.jfineco.
77 (4): 83–103. doi:10.1080/0015198X.2021.1954377. 2007.02.006.
Kelly, B., D. Palhares, and S. Pruitt. 2023. “Modeling Corporate Newey, W. K., and K. D. West. 1987. “Hypothesis Testing with
Bond Returns.” The Journal of Finance (4) 78: 1967–2008. doi: Efficient Method of Moments Estimation.” International
10.1111/jofi.13233. Economic Review 28 (3): 777–787. doi:10.2307/2526578.
Leippold, M., Q. Wang, and W. Zhou. 2021. “Machine Learning Rasekhschaffe, C., and R. Jones. 2019. “Machine Learning for
in the Chinese Stock Market.” Journal of Financial Economics Stock Selection.” Financial Analysts Journal 75 (3): 70–88. doi:
145 (2): 64–82. doi:10.1016/j.jfineco.2021.08.017. 10.1080/0015198X.2019.1596678.
Lesmond, D., J. Ogden, and C. Trzcinka. 1999. “A New Reichenbacher, M., P. Schuster, and M. Uhrig-Homburg. 2020.
Estimate of Transaction Costs.” Review of Financial Studies 12
Expected Bond Liquidity. Working Paper.
(5): 1113–1141. doi:10.1093/rfs/12.5.1113.
Roll, R. 1984. “A Simple Implicit Measure of the Effective Bid-
Leung, E., H. Lohre, S. Mischlich, Y. Shea, and M. Stroh. 2021.
Ask Spread in an Efficient Market.” The Journal of Finance (4)
“The Promises and Pitfalls of Machine Learning for Predicting
39: 1127–1139. doi:10.2307/2327617.
Stock Returns.” The Journal of Financial Data Science 3 (2): 21–
50. doi:10.3905/jfds.2021.1.062. Sarr, A., and T. Lybek. 2002. Measuring Liquidity in Financial
Lewellen, J. 2015. “The Cross-Section of Expected Stock Markets. IMF Working Paper 2002/232.
Returns.” Critical Finance Review 4 (1): 1–44. doi:10.1561/104. Tobek, O., and M. Hronec. 2021. “Does It Pay to Follow
00000024. Anomalies Research? Machine Learning Approach with
L'ubos, P., and R. F. Stambaugh. 1999. “Costs of Equity Capital International Evidence.” Journal of Financial Markets 56:
and Model Mispricing.” The Journal of Finance 54 (1): 67–121. 100588. doi:10.1016/j.finmar.2020.100588.
doi:10.1111/0022-1082.00099.
Warga, A. 1992. “Bond Returns, Liquidity, and Missing Data.”
Mahanti, S., A. Nashikkar, M. Subrahmanyam, G. Chacko, and The Journal of Financial and Quantitative Analysis (4) 27: 605–
G. Mallik. 2008. “Latent Liquidity: A New Measure of 617. doi:10.2307/2331143.

Volume 80, Number 3 127