Professional Documents
Culture Documents
SSRN Id4726993
SSRN Id4726993
ABSTRACT
In recent decades, machine learning (ML) algorithms has gained wide popularity in the
finance literature. The goal of this research is to exploit machine learning techniques in order
to analyze the effect of exchange-traded fund (ETF) illiquidity on tracking errors. We
demonstrate the superior performance of the machine learning models – Random Forest and
Gradient Boosting Decision Tree, in particular - over traditional linear models in predicting
U.S. ETF’s tracking errors. Moreover, our variable importance analysis suggests that the
features such as underlying assets based on U.S. assets (Invested in US Asset) and expense
ratio (Expense Ratio), are two key factors in the determination of predicting the tracking
errors on the ETF illiquidity. Finally, we further conduct SHAP (Shapley Additive
exPlanations) technique in order to observe the impact of a particular variable(feature) on
the difference between the considered- and average predictions of our machine learning
models. Our results indicate that the most relevant variable is Invested in US Asset, which is
in align with the previous importance analysis.
What are the determinants in the effect of exchange-traded fund (ETF) liquidity on tracking
errors? Despite the growing number of the researches on ETF, this question remains largely
unexplored. In past years, substantial number of researches paid attention to the
performance aspects of ETF (Dorocáková, 2017; Elton et al., 2019; Gallagher & Segara, 2005;
Gastineau et al., 2004; Harper et al., 2006; Levy and Lieberman, 2013; Lu et al., 2009;
Marshall et al., 2013; Svetina, 2010; Wu et al., 2021), volatility and risk premium (Ben-David
et al., 2018; Hurlin et al., 2019), the relationship with underlying portfolio (Box et al., 2021)
and so on. On contrary, the studies on tracking error which arise from the illiquidity aspect
of ETF is still quite limited (Bae and Kim, 2020; Ben-David et al., 2017; Broman and Shum,
2018; Dorocáková, 2017; Madhavan, 2014; Pu and Xie, 2022). We further delve into the topic
by employing machine learning methodologies which recently gained wide popularity in
finance researches.
One of the well-known studies on illiquidity of ETF listed in U.S. market (U.S. ETF,
hereafter) and its tracking errors, is Bae and Kim (2020). Specifically, they discuss the effect
of U.S. illiquidity on tracking errors and illustrate whether ‘positive liquidity premium’
exists in U.S. ETF market. We set their rsearch as theoretical framework upon which our
machine learning techniques are tested. Our study further captures the determinants of
tracking errors which are the results of illiquidity of U.S. ETF, and present the interaction
between the variables by conducting SHAP (Shapley Additive exPlanations) tests. To our
best knowledge, this is the first study to employ machine learning methodologies on the
illiquidity of U.S. ETF and tracking errors. We further document that the tracking error of
U.S. ETF is best captured and explained by the variables including underlying portfolio
based on U.S. market (Invested in US Asset) and expense ratio (Expense Ratio), which is in
align with previous researches on ETF (Aber et al., 2009; Dorocáková et al., 2017; Johnson,
2009; Saunders, 2018; Tsalikis & Papadopoulos, 2019).
The U.S. ETF market provides an ideal laboratory to examine the tracking error of ETF
which is mainly the result of its illiquidity (Buetow & Henderson, 2012; Chen et al., 2017;
Guo & Leung, 2015; Osterhoff & Kaserer, 2016; Shin & Soydemir, 2010). In the fourth quarter
of 2022, average daily trading volumes for U.S. equities and U.S. ETFs were $502.2 billion
and $157.7 billion, respectively. This means that U.S. ETFs accounted for 31.4% of the total
However, our paper suggests that this “low liquidity” of the less popular ETFs, does not
only attribute to relatively low trading volume, but also to other factors which may relate to
the characteristics of the ETFs and a financial market. For example, if an investor chooses to
invest in two ETFs with similar liquidity, some factors such as underlying portfolio (e.g.,
U.S. invested, emerging market, currency), or dollar trading volume (DTV) could be of more
concern to him or her. Further, we depart from previous researches in a way that we focus
more on the “explanatory characteristics” of the determinants and their interaction in
explaining the tracking errors of U.S. ETFs, which could be unique to the U.S. financial
market. While majority of previous studies on the effect of U.S. illiquidity on tracking errors
– for example, Bae and Kim (2020) – concentrate on explaining secondary market liquidity
issue.
As for a passive ETF, popular structures of ETF are physical ETF and synthetic ETF. While
a physical ETF tracks the index targeted by holding the underlying asset of index, a
synthetic ETF replicates the index by using derivatives. The advantage of synthetic ETF is its
cost-efficiency, in contrast with physical ETF which is quite expensive as it tracks emerging
market or less liquid market index. Also, as for active funds, it is widely known that the
tracking error of ETF relate to the capability of ETF managers. If the ETF managers record
high tracking errors in replicating the index, it is believed that they underperform the index
ETF. In this aspect, it is rational to believe that if ETF managers are “capable”, they would be
able to handle with low liquidity in ETF and outperform the market to some extent (Koont et
al., 2022; Li, 2022).
Following Bae and Kim (2020), we exclude the data of active funds in our dataset as the
tracking errors of active funds can be the results of “management style”. By focusing on
passive ETF in analysis, we could better capture the determinants in the effect of U.S. ETF’s
illiquidity on tracking errors through machine learning techniques. To be specific, we
believe that if the prediction performance of our machine learning methodologies is superior
to those of traditional methodologies, it would be rational to assume that the machine
1 Global ETF Market Facts: Three Things to Know from Q4 2022, 2023, Nasdaq.
hZps://www.nasdaq.com/articles/global-etf-market-facts:-three-things-to-know-from-q4-
2022#:~:text=In%20the%20fourth%20quarter%20of%202022%2C%20average%20daily,volume%20in%20the%20secondary%20m
arket%20over%20the%20quarter.
2 See Table 1 in Appendix for further details.
The major findings of our paper are following. First, we demonstrate the superior
performance of machine learning methodologies (e.g., Random Forest, GBDT) over
traditional regressions (e.g., Linear regression, LASSO, Ridge). This implies that machine
learning successfully handles with nonlinearities and overfitting issues which are embedded
in U.S. ETF dataset. Second, our machine learning analysis reveals that underlying assets
based on U.S. assets (Invested in US Asset) and expense ratio (Expense Ratio) are two key
determinants which explain the effect of illiquidity of ETF listed in U.S. on their tracking
error. This result well aligns with the traditional view that investors prefer cost-efficiency
and low management fee. To be specific, in economic downturn, investors could be more
likely to apportion underlying portfolio into categories of developed market (e.g., U.S.,
Europe) over developing market (Jhunjhunwala & Sethi, 2022; Madhavan & Maheswaran,
2016; Madhavan, 2014; Narend & Thenmozhi, 2019; Sarkar et al., 2013). Third, our SHAP
analysis indicates that the aforementioned determinants, especially Invested in US Asset,
tend to negatively affect the model output, which implies that these determinants decrease
the tendency of outputs (tracking error).
The contributions of this research are as follows. First, to our knowledge, this study is first
research to adopt machine learning techniques in analyzing the illiquidity effect of U.S. ETF.
Thanks to the superior performance of machine learning, we verify the determinants -
Invested in US Asset and Expense Ratio, in particular – in the effect of U.S. ETF’s illiquidity on
their tracking errors, which aligns with previous researches (Blitz & Huij, 2012; Charupat &
Miu, 2013; Shin & Soydemir, 2010; Tsalikis & Papadopoulos, 2019; Zawadzki, 2020). Second,
our findings reaffirms that the importance of specific variables in predicting the tracking
errors which come from illiquidity of U.S. ETFs. The “consistence” in our machine learning
models in capturing Invested in US Asset, suggest that its importance could be largely
underestimated which could be due to previous researches’ overemphasizing trading
volume as indicator of the tracking error (Dorocáková, 2017; Gallagher & Segara, 2005; Shin
& Soydemir, 2010; Tsalikis & Papadopoulos, 2019; Vardharaj et al., 2004). This is also
explained by our analysis of traditional methodologies, which suggest Log(AUM) as main
determinant.
The rest of our paper is constructed as following. Section 2 discusses previous researches
and Section 3 presents data and methodology. Section 4 introduces analysis and, followed
by discussion in Section 4. Finally, Section 5 concludes.
Generally, ETFs tend to co-move with their respective underlying portfolio, which may
pose systematic (in)stability as investors may “simultaneously” benefit from ETF’s return or
face losses. In an effort to deliver stable return, a number of ETF has grown rapidly along
with the wider range of underlying portfolios such as bonds, commodities, currencies and
equities. Particularly, previous researches indicate that due to the popularity of top ETFs,
relative non-popular ETFs suffer from low liquidity, which causes high transaction costs and
tracking errors, authorized participants (AP) could be discouraged to replicate underlying
portfolio (Bae and Kim, 2020; Dorocáková, 2017; Pu and Xie, 2022)
By definition, tracking error of ETF refers to the standard deviation of the difference in
return of ETF portfolio and the underlying portfolio's returns (Vardharaj et al., 2004). It is
widely believed that the occurrence of tracking error is not avoidable, so that APs endeavor
to lower tracking errors as low as possible. If a decoupling of returns of ETF and its
underlying portfolio widens, it could lead to the loss of faith in the liquidity formation of
ETFs. In this context, previous researches note that the fundamental to the performance of
ETF are price volatility of underlying portfolio and their arbitrage mechanism (Ben-David et
al., 2018; Grossman, 1987; Humphreys & McClain, 1998; Pagano et al., 2019; Tuzun, 2013). At
the same time, it is believed that trading ETFs could also lead to price discovery for
underlying assets, which are mainly results of the interaction between liquidity of ETFs and
“noisy” traders (Box et al., 2021; Ivanov et al., 2013).
Previous researches note that the appearance of tracking error may attribute to a number of
factors. First, the tracking errors mostly appear on the basis of the regional feature of
underlying portfolios. For instance, using the samples of different ETFs issued by iShares,
six for each of three regions (two Americas, two Asia and two Europe), Zawadzki (2020)
note that ETFs fail to mimic their underlying indexes by showing that calculated tracking
errors are often negative significantly and even further, the value of tracking errors depends
on the region and market maturity. Second, it is also notable that economic events influence
the tracking error of returns of ETF. For example, Svetina (2010) demonstrates that the 2008-
2009 financial crises negative affects the daily tracking performance for all ETF listed on the
London Stock Exchange (LSE).
Lastly, most importantly, our research is closely linked with Bae and Kim (2020), who
document that when ETFs are not liquid, tracking errors are large. Their research
demonstrates that the illiquidity of ETF relates to ETF tracking errors along with variance
and expected returns. They further imply that because of ETF illiquidity, APS increase the
costs for transaction costs for arbitrage trading, since they fail to immediately address
tracking errors, which means their incapability to properly track the underlying index. As
well, it is worthwhile to note that ETFs structure also relate to the effect of ETF illiquidity on
As aforementioned, we set the research framework of Bae and Kim (2020) as baseline and
selectively employ their variables. For example, their category of control variables segment
into (1) size and volatility, (2) characteristics of ETF and (3) synthesis and management
expense. First, Log(AUM), Log(Dollar Trading Volume), Index Volatility, Log(Shares Outstanding)
and Shares Volatility belong to first category, while Equity-Type ETF, Invested in US Assets,
Swap Based, Derivates Based, Leveraged Fund, Futures Available and Options Available are
included in second category. The variables of third category are In-Kind, Optimized, and
Expense Ratio. We download financial data composing each variable from Bloomberg and
NYSE Trade and Quote (TAQ) database. The details and sources of our variables are
presented in following the Table 3.
Next, we present the measure of ETF illiquidity. We borrow the liquidity measure from Bae
and Kim (2020) by presenting the daily relative effective half-spread which is the absolute
difference between the price of quote midpoint and its trade price. Similar to our
benchmarking study, we calculate data from NYSE TAQ database. The TAQ database has
daily-based intraday transaction data – both trades and quotes – for securities listed on
NYSE, American Stock Exchange (AMEX) and Nasdaq National Market System (NMS). In
the eq. (1) below, each security’s daily relative effective half-spread is standardized by
traded price, and is summed together to obtain average value of effective half-spread. In the
%!"
$ $
1 |𝑝!,# − 𝑚!,# |
𝑐#$ = $ ) $
⋯ (1)
𝑛# 𝑝!,#
!&'
Subsequently, we present two equations for tracking error measures, which are daily-based
and yearly-based respectively. First, the eq. (2) and (3) present daily tracking error measure
for the return difference between ETF and Net Asset Value (or Index). We calculate daily
tracking errors for two return difference (ETF returns – NAV returns, ETF returns – Index
returns) which are in the form of absolute value. Second, the eq. (4) shows yearly tracking
errors which are defined as the absolute value of the difference between one and the
coefficient of 𝑋 from the regression of 𝑌 on 𝑋. In each equation, 𝑟!,# , 𝑣# , 𝑓# and 𝛽!,# denote
the daily ETF, net asset value (NAV), returns of index (IND) and coefficient of the
regression, respectively. Subsequently, the summary statistics and correlation of our
measures of tracking errors are presented in the Table 4.
Next, we present brief explanations for traditional and machine learning methodologies for
our analysis. As aforementioned, we employ traditional methodologies such as
(multivariate) linear regression, LASSO and Ridge, and machine learning algorithms
including Random Forest and GBDT (Gradient Boosted Decision Tree).
Now we explain our machine learning algorithms which are Random Forest and GBDT.
First introduced by Breiman (2001), a random forest is an estimator in machine learning
algorithms which fits classifying decision tress on a variety of randomly selected subsamples
in dataset. Then, the algorithm averages them to improve the accuracy of the prediction, and
at the same time, control any over-fits. As well, the other machine learning algorithm,
GBDT, fits an approximation, 𝑦"! , of the function 𝐹(𝑥! ), which maps set of 𝑥 to the values
of output in a stage-wise fashion. In forms of the decision trees, the algorithm allows the
optimization of arbitrary differentiable loss function (Bentéjac et al., 2021)
Now, we present the results of empirical analysis. We first conduct performance analysis
with different methodologies, and compare the performance of prediction among them. The
Table 5 and 6 blow show the results of training and test set for the prediction of two daily
based tracking error measures (| ETF-IND|, |ETF-NAV|) and two yearly based tracking
error measures (θ(ETF-IND), θ(ETF-NAV)) respectively. In both tables, the performance of
prediction is measured by R² and four widely used metrics of regression - Mean Absolute
Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and Median
Absolute Error Loss (MedianAE).
First, in the Table 5, for daily-based tracking errors, it seems clear that the prediction
performance of R² for machine learning algorithm is similar to the traditional regression for
| ETF-IND|; however, the performance of R² is much larger for |ETF-NAV|, suggesting that
the performance for daily-based tracking errors depends on the use of the measures. On the
other hand, for the yearly-based tracking errors, the prediction performance of R² for
machine learning algorithm is much larger for both measures of tracking errors, which are
θ(ETF-IND), θ(ETF-NAV). Also, most of our measures for the prediction errors (MAE, MSE,
RMSE and MedianAE) indicate that the prediction errors of machine learning algorithms for
yearly-based tracking errors (θ(ETF-IND), θ(ETF-NAV)) are much smaller than those of
traditional methodologies. In contrast, there seems to be not very much difference between
the machine learning algorithms and traditional methods in the prediction values for daily
based tracking errors (| ETF-IND|, |ETF-NAV|).
Now, most importantly, we turn our attention to the variable importance analysis for
traditional methodologies (Linear, LASSO and Ridge) and machine learning algorithms
(Random Forest and GBDT), which is presented the Table 7. First, the results for the
traditional methods point out Log(AUM) as the most determinant in the US ETF’s illiquidity
effect on tracking errors, regardless of frequency (daily, yearly) and the types of the
methods.
On the other hand, however, the results for machine learning algorithms are quite
constrasting between daily and yearly measures for tracking errors. The results for daily-
based tracking errors are quite mixed for both trained and test set. For example, the first
measure (| ETF-IND|) suggests Intraday Volatility as the most determinant for both Random
Forest and GBDT for training set; however, the same measure for the test set presents Lev.
Fund and Log(AUM) as the most determinant respectively. The “inconsistency” in the results
is quite similar for the second measure (|ETF-NAV|). The analysis suggests Shares Outst. as
the most determinant for both Random Forest and GBDT for training set, but the test set
indicates that Log(AUM) and Log(Shares outst.) are the most determinant respectively.
Interestingly, however, the results for machine learning algorithms are consistent for
yearly-based tracking error measures (θ(ETF-IND), θ(ETF-NAV)). For instance, the results
for both measures of Random Forest and GBDT indicate that Inv. US asset is the most
determinant, followed by Expense Ratio. The only difference is that while Random Forest
analysis for first measure (| ETF-IND|) mostly indicates Log(DTV) as the third determinant
variable, the GBDT analysis consistently suggest Index Volatility as the third determinant,
regardless of the types of yearly measures.
In summary, in contrast with traditional methodologies which point out “size aspect
(Log(AUM))” of US ETFs’ tracking error, regardless of the frequency of measures, the
machine learning algorithms excels in its prediction capability, especially for the yearly-
based tracking error measures (θ(ETF-IND), θ(ETF-NAV)). The rational for the machine
10
We now conduct SHAP technique to provide the explain-ability of our model. Introduced
by L. Shapley, the Shapley value originated from the concept and method of game theory.
For example, the value of each feature is a “player” in game where the payoff is our
prediction. In a collation game (𝑁, v), where 𝑁 and v correspond to a set of players and
values respectively, the Shapley value divides the total payoff, v(𝑁) in order to estimate the
impact of the value of a particular feature on the difference between the considered- and
average prediction, given the current values of all of the features (Lundberg & Lee, 2017). On
the basis of Symmetry, Linearity and Dummy player, the difference from the average
prediction is distributed fairly among the feature values. Figure 1 and Figure 2 below
presents the daily-based and yearly-based SHAP beeswarm plots for SHAP values
respectively.
As evident in Figure 1, all of the panels fail to capture the relevance aspects of the
determinants in our analysis. This could be due to “noisy” characteristics of daily short-term
movements of passive funds, which our machine learning algorithms do not adequately
explain.
In contrast, all of the panels in Figure 1 shows that the variable which has the greatest
relevance appears to be Invested in US Asset, which is the most determinant in the previous
variable importance analysis. To explain further, this variable is mostly high with a negative
sign of SHAP value, suggesting that higher counts of Invested in US Asset tend to negatively
affect the outcome, which is tracking errors from U.S. ETF illiquidity. On the other hand, the
11
V. Discussion
Our research employs a variety of approaches explaining the tracking errors from U.S. ETF
illiquidity. Apart from Bae and Kim (2020), we attempted to capture the most determinant in
capturing the tracking errors by adopting machine learning algorithms – Random Forest
and GBDT, namely -, which showed superior performance to the traditional methodologies.
Besides the prediction performance of the machine learning algorithms, there are a number
of points to be discussed.
First, capturing Invested in US Asset as the most determinant by the machine learning
techniques suggest that this feature could have been largely underestimated in the literature
on tracking errors of ETFs. For example, the feature related to “size effect” – number of
stocks, fund size, for example – has traditionally been emphasized by previous literatures as
a variable which affects the tracking errors of ETF (Dorocáková, 2017; Vardharaj, R. et al.,
2004). This is also confirmed by our traditional methods, which consistently suggests
Log(AUM) as a main determinant. When the lack of liquidity is concerned, however, both of
our machine learning algorithms point out Invested in US Asset as the most determinant,
while the traditional methods still present Log(AUM) as the most determinant. This outcome
suggests that the investors’ tendency to follow “safe havens” such as long-term Treasuries or
the dominant US dollar could be reflected in actual outcomes (tracking errors) of US ETF
illiquidity (Habib et al., 2020; Kaczmarek et al., 2022; Kim, 2021). Our Shapley value analysis
further confirms the reliability of the outcomes of the variable importance test.
Second, our machine learning analysis is still in align with traditional random walk theory
(Cootner, 1964; Fama, 1965) in that only “yearly-based” analysis on the tracking errors of US
ETF illiquidity provides reliable outcomes, when compared with “daily-based” outcomes.
For example, the consistent results in terms of prediction performance and capturing a
determinant (Log(AUM)) are only presented in yearly-based outcomes in Table 5 through 7
and both Figure 1 and 2. The possible explanation is that the adoption of machine learning
algorithms do not significantly alter previous findings that short-term movement in future
movements is almost impossible to be predicted in advance. As aforementioned, this could
be due to the “noisy” characteristics of short-term movement of funds. Our explanation is
also in align with previous findings that employed machine learning algorithms on ETF
analysis, which mainly point out the prediction of its “yearly” performance (Liew and
Mayster, 2017; Zheng, 2021).
12
The findings of our research are following. First, the performance of machine learning
methodologies (e.g., Random Forest, GBDT) is superior to traditional regressions (e.g.,
Linear regression, LASSO, Ridge) in forecasting the errors of U.S. ETF illiquidity on tracking
errors. This suggests that machine learning algorithms better address nonlinearities and
overfitting issues embedded in dataset. Second, our study reveals that underlying assets
based on U.S. assets (Invested in US Asset) and expense ratio (Expense Ratio) are key
determinants which play a significant role in the illiquidity of ETF listed in U.S. on their
tracking error. This result aligns with the traditional view that investors prefer cost-
efficiency and low management fee. Third, our further SHAP analysis shows that the
aforementioned determinants, especially Invested in US Asset, negatively affect the model
output, implying that it decreases the tendency of outputs (tracking error).
We contribute to the realm of researches on ETF in following ways. First, we believe that
this research is first to employ machine learning algorithms in analyzing the U.S. ETF’s
illiquidity effect on tracking errors. The superior performance of machine learning verifies
the determinants - Invested in US Asset and Expense Ratio, for example – which play role in
the relationship between U.S. ETF’s illiquidity on their tracking errors. Second, our findings
confirm the importance of the determinants – for example, Invested in US Asset – in that they
could be could be largely underestimated which could be due to previous researches’
overemphasizing trading volume as indicator of the tracking error (Dorocáková, 2017;
Gallagher & Segara, 2005; Shin & Soydemir, 2010; Tsalikis & Papadopoulos, 2019; Vardharaj
et al., 2004). As pointed out, this tendency is captured by our traditional methodologies,
which present Log(AUM) as a main determinant.
The limitations of our study are as following. Despite the adoption of Random Forest and
GBDT models, we do not take other machine learning models and deep learning models
into considerations. Further, in order to further elaborate on the meanings and implications
of the most determinant in machine learning algorithms, Invested in US Asset, further
robustness tests, such as comparative analysis with active funds, would be necessary. This
research area is left for future study.
13
Aber, J. W., Li, D., Can, L., 2009. Price volatility and tracking ability of ETFs. Journal of Asset
Management 10, 210-221.
Amini, S., Elmore, R., Öztekin, Ö., Strauss, J., 2021. Can machines learn capital structure
dynamics? Journal of Corporate Finance 70, 102073.
Angel, J. J., Broms, T. J., Gastineau, G. L., 2016. ETF transaction costs are often higher than
investors realize. The Journal of Portfolio Management 42(3), 65-75.
Avramov, D., Li, M., Wang, H., 2021. Predicting corporate policies using downside risk: A
machine learning approach. Journal of Empirical Finance 63, 1-26.
Bae, K., Kim, D., 2020. Liquidity risk and exchange-traded fund returns, variances, and
tracking errors. Journal of Financial Economics 138(1), 222-253.
Bali, T. G., Goyal, A., Huang, D., Jiang, F., Wen, Q., 2020. Predicting corporate bond returns:
Merton meets machine learning. Georgetown McDonough School of Business Research Paper
(3686164), 20-110.
Ben-David, I., Franzoni, F., Moussawi, R., 2018. Do ETFs Increase Volatility? The Journal of
Finance 73, 2471-2535.
Bentéjac, C., Csörgő, A. & Martínez-Muñoz, G., 2021. A comparative analysis of gradient
boosting algorithms. Artificial Intelligence Review volume 54, 1937–1967
Bilgin, R., 2023. The Selection of Control Variables In Capital Structure Research With
Machine Learning.
Blitz, D., Huij, J., 2012. Evaluating the performance of global emerging markets equity
exchange-traded funds. Emerging markets review 13(2), 149-158.
Broman, M.S., Shum, P., 2018. Relative Liquidity, Fund Flows and Short-Term Demand:
Evidence from Exchange-Traded Funds. Financial Review 53, 87-115.
Box, T., Davis, R., Evans, R., Lynch, A., 2021. Intraday arbitrage between ETFs and their
underlying portfolios. Journal of Financial Economics 141(3), 1078-1095.
Chen, J., Chen, Y., Frijns, B., 2017. Evaluating the tracking performance and tracking error of
New Zealand exchange traded funds. Pacific Accounting Review 29(3), 443-462.
14
Charupat, N., Miu, P., 2013. Recent developments in Exchange-Traded Fund literature:
Pricing efficiency, tracking ability, and effects on underlying securities. Managerial Finance
39(5), 427-443.
Cootner, P.H. (Ed.), 1967. The Random Character of Stock Market Prices. The MIT Press.
Dorocáková, M., 2017. Comparison of ETF´ s performance related to the tracking error.
Journal of International Studies 10(4), 154-165.
Elton, E.J., Gruber, M.J., de Souza, A., 2019. Passive mutual funds and ETFs: Performance and
comparison. Journal of Banking & Finance 106, 265-275.
Eliasy, A., Przychodzen, J., 2020. The role of AI in capital structure to enhance corporate
funding strategies. Array 6, 100017.
Fama, E.F., 1965. Random Walks in Stock Market Prices. Financial Analysts Journal, 21(5), 55-
59. Published By: Taylor & Francis, Ltd.
Gallagher, D. R., Segara, R., 2005. The performance and trading characteristics of exchange-
traded funds. Journal of Investment Strategy 1(1), 47-58.
Gastineau, G. L., 2004. The benchmark index ETF performance problem. The Journal of
Portfolio Management 30(2), 96-103.
Grossman, S. J., 1987. An analysis of the implications for stock and futures price volatility of
program trading and dynamic hedging strategies.
Guo, K., Leung, T., 2015. Understanding the tracking errors of commodity leveraged ETFs.
Springer New York.
Habib, M. M., Stracca, L., Venditti, F., 2020. The fundamentals of safe assets. Journal of
International Money and Finance, 102.
Harper, J. T., Madura, J., Schnusenberg, O., 2006. Performance comparison between exchange-
traded funds and closed-end country funds. Journal of International Financial Markets,
Institutions and Money 16(2), 104-122.
Hoerl, A.E. & Kennard, R.W., 1970. Ridge Regression: Biased Estimation for Nonorthogonal
Problems. Technometrics, 12(1), 55-67. Publisher: Taylor & Francis. doi:
10.1080/00401706.1970.10488634.
Houweling, P., 2012. On the performance of fixed-income exchange-traded funds. The Journal
of Beta Investment Strategies 3(1), 39-44.
15
Hurlin, C., Iseli, G., Pérignon, C., Yeung, S., 2019. The counterparty risk exposure of ETF
investors. Journal of Banking & Finance 102, 215-230.
Hu, Z., Zhao, Y., Khushi, M., 2021. A survey of forex and stock price prediction using deep
learning. Applied System Innovation 4(1), 9.
Ivanov, S.I., Jones, F.J., Zaima, J.K., 2013. Analysis of DJIA, S&P 500, S&P 400, NASDAQ 100
and Russell 2000 ETFs and their influence on price discovery. Global Finance Journal 24(3),
171-187.
Ivașcu, C. F., 2023. Understanding Dividend Puzzle Using Machine Learning. Computational
Economics 1-19.
Johnson, B., Bioy, H., Boyadzhiev, D., 2016. Assessing the true cost of strategic-beta ETFs. The
Journal of Index Investing 7(1), 35.
Johnson, W. F., 2009. Tracking errors of exchange traded funds. Journal of Asset Management
10, 253-262.
Jhunjhunwala, S., Sethi, A., 2022. Do ETFs affect the return co-movement of their underlying
assets? Evidence from an emerging market. Managerial Finance 48(11), 1661-1686.
Kaczmarek, T., Będowska-Sójka, B., Grobelny, P., Perez, K., 2022. False Safe Haven Assets:
Evidence From the Target Volatility Strategy Based on Recurrent Neural Network. Research
in International Business and Finance, 60, 101610. ISSN 0275-5319.
Kamalov, F., Gurrib, I., Rajab, K., 2021. Financial forecasting with machine learning: price vs
return. Journal of Computer Science 17(3), 251-264.
Kim, J. M., Kim, D. H., Jung, H., 2021. Applications of machine learning for corporate bond
yield spread forecasting. The North American Journal of Economics and Finance 58, 101540.
Kim, M., 2021. Adaptive trading system integrating machine learning and back-testing:
Korean bond market case. Expert Systems with Applications 176, 114767.
Kim, T., 2021. Safe Asset Demand, Global Capital Flows and Wealth Concentration. IMF
Working Paper No. 2021/254. Available at SSRN: https://ssrn.com/abstract=4026482.
Koont, N., Ma, Y., Pastor, L., Zeng, Y., 2022. Steering a ship in illiquid waters: Active
management of passive funds. National Bureau of Economic Research.
Levy, A., Lieberman, O., 2013. Overreaction of country ETFs to US market returns: Intraday
vs. daily horizons and the role of synchronized trading. Journal of Banking & Finance 37(5),
1412-1421.
16
Liew, J.K.S., & Mayster, B., 2017. Forecasting ETFs with Machine Learning Algorithms. Johns
Hopkins Carey Business School. Version 1.3. http://etfprediction.pythonanywhere.com/. Date:
January 14, 2017.
Lu, L., Wang, J., Zhang, G., 2009. Long term performance of leveraged ETFs. Available at
SSRN 1344133.
Lundberg, S. M., & Lee, S.-I., 2017). A Unified Approach to Interpreting Model Predictions.
Advances in Neural Information Processing Systems (Vol. 30).
Madhavan, V., Maheswaran, S., 2016. Indian exchange traded funds: relationship with
underlying indices. Economic and Political Weekly 142-148.
Madhavan, V., 2014. Investigating the nature of nonlinearity in indian exchange traded funds
(ETFs). Managerial Finance 40(4), 395-415.
Marshall, B.R., Nguyen, N.H., Visaltanachoti, N., 2013. ETF arbitrage: Intraday evidence.
Journal of Banking & Finance 37(9), 3486-3498.
Narend, S., Thenmozhi, M., 2019. Do country ETFs influence foreign stock market index?
Evidence from India ETFs. Journal of Emerging Market Finance 18(1_suppl), S59-S86.
Obthong, M., Tantisantiwong, N., Jeamwatthanachai, W., Wills, G., 2020. A survey on
machine learning for stock price prediction: Algorithms and techniques.
Osterhoff, F., Kaserer, C., 2016. Determinants of tracking error in German ETFs–The role of
market liquidity. Managerial Finance 42(5), 417-437.
Pagano, M., Sánchez Serrano, A., Zechner, J., 2019. Can EFTs Contribute to Systemic Risk?
ESRB: Advisory Scientific Committee Reports 9.
Rompotis, G.G., 2011. The Performance of Actively Managed Exchange-Traded Funds. The
Journal of Index Investing 1(4), 53-65.
Saunders, K. T., 2018. Analysis of international ETF tracking error in country-specific funds.
Atlantic Economic Journal 46, 151-160.
Sarkar, S. S., Dutta, S., Dutta, P., 2013. A review of Indian index funds. Global Business
Review 14(1), 89-98.
Shin, S., Soydemir, G., 2010. Exchange-traded funds, persistence in tracking errors and
information dissemination. Journal of Multinational Financial Management 20(4-5), 214-234.
Svetina, M., 2010. Exchange traded funds: Performance and competition. Journal of Applied
Finance (Formerly Financial Practice and Education) 20(2).
17
Tokat, E., Hayrullahoğlu, A. C., 2022. Pairs trading: is it applicable to exchange-traded funds?.
Borsa Istanbul Review 22(4), 743-751.
Tsalikis, G., Papadopoulos, S., 2019. ETFs-Performance, tracking errors and their
determinants in Europe and the USA. Risk Governance & Control: Financial Markets &
Institutions 9(4).
Tuzun, Tugkan, 2013. Are Leveraged and Inverse ETFs the New Portfolio Insurers. FEDS
Working Paper No. 2013-48.
Vardharaj, R., Fabozzi, F.J., Jones, F.J., 2004. Determinants of Tracking Error for Equity
Portfolios. The Journal of Investing 13(2), 37-47.
Wu, C., Xiong, X., Gao, Y., 2021. Performance comparisons between ETFs and traditional
index funds: Evidence from China. Finance Research Letters 40, 101740.
Yaseen, H., Dragotă, V., 2021. Forecasting the dividend policy using machine learning
approach: Decision tree regression models. In Eurasian Business and Economics Perspectives:
Proceedings of the 31st Eurasia Business and Economics Society Conference (pp. 19-39).
Cham: Springer International Publishing.
Zawadzki, K., 2020. The performance of ETFs on developed and emerging markets with
consideration of regional diversity. Quantitative Finance and Economics 4(3), 515-525.
Zhang, C., Amir Sjarif, N.N., & Ibrahim, R., 2023. Deep learning models for price forecasting
of financial time series: A review of recent advancements: 2020–2022. WIREs Data Mining and
Knowledge Discovery.
18
19
3 Please note that the set of our control variables is basically same to Bae and Kim (2020).
20
21
22
23
24
25
26