Journal of Economic Dynamics & Control 159 (2024) 104804

Contents lists available at ScienceDirect

Journal of Economic Dynamics and Control

journal homepage: www.elsevier.com/locate/jedc

A contagion test with unspecified heteroscedastic errors

Ernest Aboagye a , Stanley Iat-Meng Ko b , Chia Chun Lo c,d , Cody Yu-Ling Hsiao e,∗ ,
Liang Peng a
a Maurice R. Greenberg School of Risk Science, Georgia State University, United States of America
b Graduate School of Economics and Management, Tohoku University, Japan
c Huatai Futures Co., Ltd, China
d Doctoral Think Tank, Macau, China
e Centre for Gaming and Tourism Studies, Macao Polytechnic University, China

A R T I C L E I N F O A B S T R A C T

JEL classification: The tests of contagion in Fry-McKibbin et al. (2010) filter returns by a vector autoregressive
G12 model, assume residuals are independent, fit a parametric distribution family to residuals, and
G15 test for the change of contagion measures, which ignore the effect of filtering the time series
C51
model and the stylized fact of heteroscedasticity in daily returns. This paper studies a contagion
test based on correlation by allowing heteroscedastic errors with a deterministic jump from the
Keywords:
pre-crisis to the crisis periods. Because the developed test does not infer heteroscedasticity, it
Contagion test
Heteroscedasticity is robust against heteroscedasticity but its asymptotic variance under the null hypothesis of no
Variance estimation contagion becomes complicated, which relies on a block method for estimating the asymptotic
variance. A simulation study confirms the good finite sample performance of the new contagion
test. Finally, we apply the test to three datasets to test for contagion.

1. Introduction

It is well-acknowledged in the international finance literature that correlation tightens when an international financial crisis
takes place, see for example, Ang and Chen (2002) and Longin and Solnik (2001), among many others. Forbes and Rigobon (2002)
further examine the “extreme” market co-movements. Typically, the residuals are first extracted from the estimated conditional mean
model and the residual co-movement across markets is termed as the contagion.1 Thus, contagion is regarded as the excess market
co-movement. Forbes and Rigobon (2002) test a change in the residual correlation coefficients before and after a financial crisis.
They conclude that no contagion (excess co-movement) was found during three financial crises, namely the 1997 Asian crisis, the
1994 Mexican devaluation, and the 1987 United States (US) market crash.
Suppose we observe two financial return sequences of {𝑋𝑡,𝓁 }𝑇𝑡=1 in two different markets (𝓁 = 1, 2) and a financial crisis in one
market takes place in the period (𝑚, 𝑇 ]. A simple co-movement study compares the sample correlation coefficients “conditional” on
pre-crisis and crisis periods, but Forbes and Rigobon (2002) and Ronn et al. (2009) argue that the conditional correlation coefficient

* Corresponding author.
E-mail address: [email protected] (C.Y.-L. Hsiao).
1
Forbes and Rigobon (2002), Bouri et al. (2022) and Wu et al. (2023) use the VAR model to extract the residual correlations as indicative of contagion; while other
researchers define unexplained shocks using the factor models of Bekaert et al. (2009), Bekaert et al. (2014) and Dungey and Gajurel (2015), and the multinomial
logistic regression model of Bae et al. (2003).

https://doi.org/10.1016/j.jedc.2023.104804
Received 14 August 2023; Received in revised form 14 December 2023; Accepted 15 December 2023
Available online 21 December 2023
0165-1889/© 2023 Elsevier B.V. All rights reserved.
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

is not a good measure and propose the “unconditional” correlation coefficients. More specifically, Forbes and Rigobon (2002) assume
the linear model 𝑋𝑡,2 = 𝛼 + 𝛽𝑋𝑡,1 + 𝜀𝑡 and show the conditional correlation coefficient between 𝑋𝑡,2 and 𝑋𝑡,1 given 𝑋𝑡,1 in the crisis
will be larger than that given 𝑋𝑡,1 in the pre-crisis because 𝑋𝑡,1 has a bigger variance in the crisis than the pre-crisis, which motivates
them to test for a change in the unconditional correlation coefficient by adjusting the bias from the conditional one. On the other
hand, Ronn et al. (2009) assume a bivariate normal distribution for (𝑋𝑡,1 , 𝑋𝑡,2 ) and show that
( )1∕2
( ) 1+𝛿
Corr 𝑋𝑡,1 , 𝑋𝑡,2 |𝑋𝑡,1 is subject to crisis in (𝑚, 𝑇 ] = 𝜌, (1)
1 + 𝛿𝜌2
where 𝛿 = 𝜎𝑐,1 ∕𝜎𝑝,1 − 1, 𝜎𝑝,1 and 𝜎𝑐,1 are the volatilities in the pre-crisis and crisis periods, and 𝜌 is the original correlation in the
bivariate normal distribution of (𝑋𝑡,1 , 𝑋𝑡,2 ). Again, this motivates the use of the unconditional correlation coefficient for a contagion
test via a bias correction based on the above equation (1). See also Boyer et al. (1999) and Loretan and English (2000) for similar
discussions. Fry-McKibbin et al. (2010) extend the correlation contagion test to a higher co-moment test. Since then, the contagion
literature has developed various higher co-moment tests, see for example, single-channel tests of co-kurtosis and co-volatility (Fry-
McKibbin and Hsiao, 2018), multiple-channel tests of correlation and co-skewness (Chan et al., 2019), multiple-channel tests of
co-skewness, co-kurtosis, and co-volatility (Fry-McKibbin et al., 2019), multiple-channel tests of correlation, co-skewness, co-kurtosis,
and co-volatility (Fry-McKibbin et al., 2021; Hsiao and Morley, 2022).
Statistically, the aforementioned papers assume that the data are independent and follow a parametric distribution family after
filtering the conditional means. They test whether there is a change in the employed contagion measure between the pre-crisis and
crisis periods without considering the filtering effect. However, it is essential to address the filtering effect and heteroscedasticity
in contagion tests, given that heteroscedasticity is a common characteristic of financial returns. Consequently, the question arises
as to how to account for the filtering effect and heteroscedasticity in a contagion test. Moreover, because of heteroscedasticity, it is
extremely challenging to speculate a parametric distribution family for the data after only filtering the conditional means.
To develop a contagion test allowing heteroscedasticity and without fitting a parametric distribution family, we measure the
financial contagion between two markets by the correlation coefficient between the filtered returns 𝜂𝑡,1 = (𝑋𝑡,1 − 𝜇𝑡,1 )∕𝜎𝑡,1 and
𝜂𝑡,2 = (𝑋𝑡,1 − 𝜇𝑡,2 )∕𝜎𝑡,2 , where 𝜇𝑡,𝓁 and 𝜎𝑡,𝓁 are conditional mean and conditional standard deviation of 𝑋𝑡,𝓁 given the 𝜎 -field 𝑡−1,𝓁
generated by {𝑋𝑠,𝓁 ∶ 𝑠 ≤ 𝑡 − 1}. We assume that both {𝜂𝑡,1 }𝑇𝑡=1 and {𝜂𝑡,2 }𝑇𝑡=1 are a sequence of independent and identically distributed
random variables with mean zero and variance one but the joint distribution of (𝜂𝑡,1 , 𝜂𝑡,2 ) differs in pre-crisis and crisis when contagion
exists. To have a robust contagion test against heteroscedasticity, we propose to test for a change in the covariance of (𝑋𝑡,1 −
√ √
𝜇𝑡,1 )∕ 𝐸(𝜎𝑡,1
2 ) and (𝑋 − 𝜇 )∕ 𝐸(𝜎 2 ) between pre-crisis and crisis. Because covariance estimators in pre-crisis and crisis have
𝑡,2 𝑡,2 𝑡,2
very complicated asymptotic variances due to the use of 𝐸(𝜎𝑡,𝓁 2 ), we employ the overlapping block method in Peligrad and Shao

(1995) for 𝜌-mixing sequences and Shi (2000) for 𝛼 -mixing sequences to estimate the asymptotic variances. Alternatively, we can
employ the nonoverlapping block method in Carlstein (1986) to estimate the asymptotic variance or employ a block bootstrap
method to compute the critical values instead of estimating the asymptotic variance such as the moving block bootstrap method in
Künsch (1989). The key idea in using an overlapping or a non-overlapping block method is that many statistics computed blocks
have a weak dependence due to mixing conditions.
Although our method can be extended to other higher co-movement tests mentioned above, extra care should be taken to deal
with the heavy tail effect of financial returns, which is beyond the scope of this paper. Empirically, we examine the 2019 COVID
crisis, the 2007 subprime mortgage crisis in the US, and the 1997 Asian financial crisis, and compare our results with those using the
conventional parametric FR test in Forbes and Rigobon (2002).2
We organize the paper as follows. Section 2 presents the methodology and the asymptotic results of our contagion statistic. A
simulation study is presented in Section 3. Section 4 is some data analyses. Section 5 concludes. All theoretical derivations are put in
the Appendix.

2. A robust contagion test

We model each series of returns by an autoregressive model with unspecified heteroscedastic errors:
{ ∑𝑝𝓁
𝑋𝑡,𝓁 = 𝜇𝓁 + 𝑖=1 𝜙𝑖,𝓁 𝑋𝑡−𝑖,𝓁 + 𝜀𝑡,𝓁 , 𝜀𝑡,𝓁 = {1 + 𝜆𝓁 𝐼(𝑡 > 𝑚)}𝑒𝑡,𝓁 ,
(2)
𝑒𝑡,𝓁 = 𝜎𝑡,𝓁 𝜂𝑡,𝓁 , 𝜎𝑡,𝓁 = ℎ𝓁 (𝑒𝑡−1,𝓁 , 𝑒𝑡−2,𝓁 , ⋯)
for 𝑡 = 1, ⋯ , 𝑇 and 𝓁 = 1, 2, where 𝜆𝓁 > −1 is a deterministic volatility jump from the pre-crisis to the crisis, {𝜂𝑡,𝓁 }𝑇𝑡=1 for 𝓁 = 1 and
2 is a sequence of independent and identically distributed random variables with mean zero and variance one, {𝜼𝑡 = (𝜂𝑡,1 , 𝜂𝑡,2 )⊤ }
is a sequence of independent random vectors with the same joint distribution in case of no contagion but with two different joint
distributions in the pre-crisis period and the crisis period in case of contagion, ℎ𝓁 is a measurable function such that {𝑒𝑡,𝓁 } is strictly
stationary, and 𝑚 is known to be the last date of the pre-crisis period of one asset. Here, we allow 𝜂𝑡,1 and 𝜂𝑡,2 to be dependent with
𝐸(𝜂𝑡,1 𝜂𝑡,2 ) = 𝜌𝑡 , which is constant in pre-crisis and crisis. That is, we assume that 𝜌𝑡 ’s are equal to 𝜌𝑝,𝜂 for 𝑡 ≤ 𝑚 and 𝜌𝑐,𝜂 for 𝑡 > 𝑚.

2
Empirical studies of the impacts of the three financial crises can be found in the literature, e.g., Islam and Volkov (2022) for the Asian crisis of 1997, Kangogo et
al. (2023) for the global financial crisis of 2007, and Xu et al. (2023) for the COVID-19 pandemic, among many others.

2
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Also, the above model for {𝑒𝑡,𝓁 } is quite general and includes the well-known GARCH models and their many variants; see Zhu and
Ling (2015).
Under model (2), we measure the contagion by the correlation coefficient between 𝜂𝑡,1 and 𝜂𝑡,2 and develop the contagion test by
testing

𝐻0 ∶ 𝜌𝑝,𝜂 = 𝜌𝑐,𝜂 against 𝐻𝑎 ∶ 𝜌𝑝,𝜂 ≠ 𝜌𝑐,𝜂 . (3)

As argued in the introduction, we want a robust contagion test without modeling and estimating the heteroscedasticity, i.e., a
contagion test using estimated (𝜀𝑡,1 , 𝜀𝑡,2 )’s rather than (𝜂𝑡,1 , 𝜂𝑡,2 )’s. Note that the imposed structure on 𝜀𝑡,𝓁 ’s in (2) not only allows a
variance jump from the pre-crisis to the crisis but also implies that

⎛ ⎞ ⎛ ⎞
⎜ 𝜀𝑡,1 𝜀𝑡,2 ⎟ ⎜ 𝑒𝑡,1 𝑒𝑡,2 ⎟ 𝐸(𝜎𝑡,1 𝜎𝑡,2 )𝐸(𝜂𝑡,1 𝜂𝑡,2 )
𝐸 ⎜√ √ ⎟ = 𝐸 ⎜√ √ ⎟= √ √ .
⎜ 𝐸(𝜀2 ) 𝐸(𝜀2 ) ⎟ ⎜ 𝐸(𝑒2 ) 𝐸(𝑒2 ) ⎟ 𝐸(𝜎𝑡,1
2 ) 𝐸(𝜎 2 )
⎝ 𝑡,1 𝑡,2 ⎠ ⎝ 𝑡,1 𝑡,2 ⎠ 𝑡,2

𝐸(𝜎𝑡,1 𝜎𝑡,2 )
Because √ √ is constant for all 𝑡 = 1, ⋯ , 𝑇 under the null hypothesis of (3), we propose to test
𝐸(𝜎𝑡,1
2 ) 𝐸(𝜎 2 )
𝑡,2

𝐻0 ∶ 𝜌𝑝,𝜀 = 𝜌𝑐,𝜀 against 𝐻𝑎 ∶ 𝜌𝑝,𝜀 ≠ 𝜌𝑐,𝜀 , (4)

where
⎛ ⎞ ⎛ ⎞
⎜ 𝜀1,1 𝜀1,2 ⎟ ⎜ 𝜀𝑇 ,1 𝜀𝑇 ,2 ⎟
𝜌𝑝,𝜀 = 𝐸 ⎜ √ √ ⎟ and 𝜌 𝑐,𝜀 = 𝐸 ⎜√ √ ⎟
⎜ 𝐸(𝜀2 ) 𝐸(𝜀2 ) ⎟ ⎜ 𝐸(𝜀2 ) 𝐸(𝜀2 ) ⎟
⎝ 1,1 1,2 ⎠ ⎝ 𝑇 ,1 𝑇 ,2 ⎠

are the correlation coefficients of 𝜀𝑡,1 and 𝜀𝑡,2 in the pre-crisis and crisis as we assume they are constant in these two periods.
Therefore, we use the difference 𝜌𝜀 = 𝜌𝑝,𝜀 − 𝜌𝑐,𝜀 for a contagion test. It is clear from the arguments above that having a deterministic
jump rather than a stochastic jump in modeling heteroscedasticity plays a key role in developing our robust contagion test without
inferring the heteroscedasticity. It is possible to generalize our robust contagion test based on the correlation coefficient to a higher
co-moment, which requires a weighted inference to reduce the heavy tail effect and is beyond the scope of this paper. Although the
null hypothesis in (3) is equivalent to that in (4), the alternative hypotheses in (3) and (4) are not equivalent. Hence, it would be
more powerful to develop a test for (3) via explicitly modeling and inferring 𝜎𝑡,1 and 𝜎𝑡,2 in (2), and we will investigate in the future.
{ }𝑚 { }𝑇
𝜀𝑡,1 𝜀𝑡,2 𝜀𝑡,1 𝜀𝑡,2
Because of heteroscedasticity, √ √ and √ √ have a complicated dependence structure
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 ) 𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
𝑡=1 𝑡=𝑚+1
and the asymptotic theory for martingale differences does not apply even when 𝜼𝑡 ’s have
the same distribution in pre-crisis and
crisis. Therefore, it is challenging to quantify the uncertainty of estimates for 𝜌𝜀 if we use a different sample size from pre-crisis and
crisis. Here, we propose to use the same sample sizes from pre-crisis and crisis and pair them as a strictly stationary sequence, in
which we can employ a block method to quantify the estimation uncertainty.
Without loss of generality, we assume 𝑚 ≥ (𝑇 − 𝑚) ∶= 𝑚0 and define

𝜻 𝓁 = (𝜇𝓁 , 𝜙1,𝓁 , ⋯ , 𝜙𝑝𝓁 ,𝓁 )⊤ and 𝒁 𝑡,𝓁 = (1, 𝑋𝑡−1,𝓁 , ⋯ , 𝑋𝑡−𝑝𝓁 ,𝓁 )⊤ .

First, we compute the following quantities for 𝓁 = 1, 2
∑
𝜻̂𝑝,𝓁 = arg min𝜻 𝓁 𝑚 ⊤
𝑡=𝑚−𝑚0 +1 {𝑋𝑡,𝓁 − 𝜻 𝓁 𝒁 𝑡,𝓁 }
2
1 ∑𝑚 ⊤ −1 1 ∑ 𝑚
= {𝑚
0
𝑡=𝑚−𝑚0 +1 𝒁 𝑡,𝓁 𝒁 𝑡,𝓁 } 𝑚0 𝑡=𝑚−𝑚0 +1 𝑋𝑡,𝓁 𝒁 𝑡,𝓁 ,
∑𝑇
𝜻̂𝑐,𝓁 = arg min𝜻 𝓁 𝑡=𝑚+1 {𝑋𝑡,𝓁 − 𝜻 ⊤ 𝒁 }2
𝓁 𝑡,𝓁
1 ∑𝑇 ⊤ −1 1 ∑𝑇
= {𝑚
0
𝑡=𝑚+1 𝒁 𝑡,𝓁 𝒁 𝑡,𝓁 } 𝑚 𝑡=𝑚+1 𝑋𝑡,𝓁 𝒁 𝑡,𝓁 ,
0
⊤
𝜀̂𝑡,𝓁 = 𝑋𝑡,𝓁 − ̂
𝜻 𝑝,𝓁 𝒁 𝑡,𝓁 for 𝑡 = 𝑚 − 𝑚0 + 1, ⋯ , 𝑚,
⊤
𝜀̂𝑡,𝓁 = 𝑋𝑡,𝓁 − ̂
𝜻 𝑐,𝓁 𝒁 𝑡,𝓁 for 𝑡 = 𝑚 + 1, ⋯ , 𝑇 ,
𝑚
∑ 𝑇
1 1 ∑ 2
𝜎
̂𝑝,𝓁
2
= 𝜀̂2𝑡,𝓁 , 𝜎
̂𝑐,𝓁
2
= 𝜀̂ ,
𝑚0 𝑡=𝑚−𝑚0 +1
𝑚0 𝑡=𝑚+1 𝑡,𝓁

leading to our test statistic

𝑚
∑ 𝑇
1 𝜀̂𝑡,1 𝜀̂𝑡,2 1 ∑ 𝜀̂𝑡,1 𝜀̂𝑡,2
𝜌̂𝜀 = 𝜌̂𝑝,𝜀 − 𝜌̂𝑐,𝜀 with 𝜌̂𝑝,𝜀 = and 𝜌̂𝑐,𝜀 = .
𝑚0 𝑡=𝑚−𝑚0 +1
𝜎
̂𝑝,1 𝜎
̂𝑝,2 𝑚0 𝑡=𝑚+1 𝜎
̂𝑐,1 𝜎
̂𝑐,2

To derive the asymptotic limit of 𝜌̂𝜀 , we employ the following regularity conditions for model (2).

3
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

• C1) For 𝓁 = 1 and 2, 𝜻 𝓁,0 is an interior point in the parameter space Θ𝓁 , and for each 𝜻 𝓁 ∈ Θ𝓁 , 𝜙𝓁 (𝑧) ≠ 0 when |𝑧| ≤ 1, where
∑𝑝𝓁
𝜙𝓁 (𝑧) = 1 − 𝑖=1 𝜙𝑖,𝓁 𝑧𝑖 .
• C2) For 𝓁 = 1 and 2, {𝜂𝑡,𝓁 }𝑇𝑡=1 is a sequence of independent and identically distributed random variables with mean zero and
variance one, and {𝜼𝑡 }𝑇𝑡=1 is a sequence of independent random vectors with the same joint distribution in case of no contagion,
i.e., 𝐸(𝜂𝑡,1 𝜂𝑡,2 ) = 𝜌𝑡 is constant for all 𝑡 = 1, ⋯ , 𝑇 , but with a joint distribution in pre-crisis and another different joint distribution
in crisis in case of contagion such that 𝐸(𝜂𝑡,1 𝜂𝑡,2 ) = 𝜌𝑡 is 𝜌𝑝 for 𝑡 ≤ 𝑚 and 𝜌𝑐 for 𝑡 > 𝑚 with 𝜌𝑝 ≠ 𝜌𝑐 .
• C3) Let 𝒆𝑡 = (𝑒𝑚−𝑚0 +𝑡,1 , 𝑒𝑚+𝑡,1 , 𝑒𝑚−𝑚0 +𝑡,2 , 𝑒𝑚+𝑡,2 )⊤ for 𝑡 = 1, ⋯ , 𝑚0 and assume that {𝒆𝑡 }∞ −∞ is a strict 𝛼 -mixing sequence with
mixing coefficient

𝛼𝑘 = sup sup |𝑃 (𝐴 ∩ 𝐵) − 𝑃 (𝐴)𝑃 (𝐵)| → 0 as 𝑘 → ∞,

𝑛≥1 𝐴∈̃ −∞
𝑛 ,𝐵∈̃ ∞
𝑛+𝑘

where ̃ 𝑛𝑛+𝑘 is the 𝜎 -field generated by {𝒆𝑡 }𝑛+𝑘

𝑡=𝑛 . Further, assume 𝐸(||𝒆𝑡 ||
4+𝛿 ) < ∞ for some 𝛿 > 0 and 𝛼 ≤ 𝐷𝜌𝑘 for some 𝐷 > 0
𝑘 0
and 𝜌0 ∈ (0, 1) as 𝑘 large enough.
• C4) Assume 𝑚 ≥ 𝑚0 ∶= 𝑇 − 𝑚 and 𝑚0 → ∞ as 𝑇 → ∞.

We remark that well-known GARCH models satisfy C3). In the future, we investigate developing a contagion test via explicitly
modeling and inferring the heteroscedasticity with a weighted inference to reduce the heavy tail effect, which will allow the use of
different sample sizes in pre-crisis and crisis and be more efficient.

Theorem 1. Under conditions C1)-C4) for model (2), we have

𝑚0
√ 1 ∑ 𝑑
𝑚0 (𝜌̂𝜀 − 𝜌𝜀 ) = √ {𝑊𝑠,𝑝 − 𝑊𝑠,𝑐 } + 𝑜𝑝 (1) → 𝑁(0, 𝜎02 ),
𝑚0 𝑠=1
where
∑
∞
𝜎02 = 𝐸{(𝑊1,𝑝 − 𝑊1,𝑐 )2 } + 2 𝐸{(𝑊1,𝑝 − 𝑊1,𝑐 )(𝑊𝑖,𝑝 − 𝑊𝑖,𝑐 )},
𝑖=2
𝜀𝑚−𝑚 +𝑠,1 𝜀𝑚−𝑚 +𝑠,2 𝜀𝑚−𝑚 +𝑠,1 𝜀𝑚−𝑚 +𝑠,2
𝑊𝑠,𝑝 = {√ 0 √ 0
− 𝐸( √ 0 √ 0
)}
𝐸(𝜀2𝑚−𝑚 +𝑠,1 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,2 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,1 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,2 )
0 0 0 0
𝜀𝑚−𝑚 +𝑠,1 𝜀𝑚−𝑚 +𝑠,2
0 0
+{ √ − 1}𝐴𝑝,1 + { √ − 1}𝐴𝑝,1
𝐸(𝜀2𝑚−𝑚 +𝑠,1 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,2 )
0 0
𝜀𝑚−𝑚 +𝑠,1 𝒁⊤𝑚−𝑚0 +𝑠,1 𝜀𝑚−𝑚 +𝑠,2 𝒁⊤
𝑚−𝑚0 +𝑠,2
+√ 0 √ Γ−1
𝑝,1
𝑨𝑝,2 + √ 0 √ Γ−1 𝑨 ,
𝑝,2 𝑝,3
𝐸(𝜀2𝑚−𝑚 +𝑠,1 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,1 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,2 ) 𝐸(𝜀2𝑚−𝑚 +𝑠,2 )
0 0 0 0
𝜀 𝜀 𝜀 𝜀
𝑊𝑠,𝑐 = { √ 𝑚+𝑠,1 √ 𝑚+𝑠,2 − 𝐸( √ 𝑚+𝑠,1 √ 𝑚+𝑠,2 )}
𝐸(𝜀2𝑚+𝑠,1 ) 𝐸(𝜀2𝑚+𝑠,2 ) 𝐸(𝜀2𝑚+𝑠,1 ) 𝐸(𝜀2𝑚+𝑠,2 )
𝜀 𝜀
+{ √ 𝑚+𝑠,1 − 1}𝐴𝑐,1 + { √ 𝑚+𝑠,2 − 1}𝐴𝑐,1
𝐸(𝜀2𝑚+𝑠,1 ) 𝐸(𝜀2𝑚+𝑠,2 )
𝜀𝑚+𝑠,1 𝒁⊤
𝑚+𝑠,1 𝜀𝑚+𝑠,2 𝒁⊤
𝑚+𝑠,2
+√ √ Γ−1 𝑨𝑐,2 + √ √ Γ−1 𝑨𝑐,3 ,
𝐸(𝜀2𝑚+𝑠,1 ) 𝐸(𝜀2𝑚+𝑠,1 ) 𝑐,1 𝐸(𝜀2𝑚+𝑠,2 ) 𝐸(𝜀2𝑚+𝑠,2 ) 𝑐,2

𝒁 𝑚,𝓁 𝒁⊤
𝑚,𝓁 𝒁 𝑚+1,𝓁 𝒁⊤
𝑚+1,𝓁
Γ𝑝,𝓁 = 𝐸{ √ √ }, Γ𝑐,𝓁 = 𝐸{ √ √ } for 𝓁 = 1, 2,
𝐸(𝜀𝑚,𝓁 ) 𝐸(𝜀𝑚,𝓁 )
2 2 𝐸(𝜀𝑚+1,𝓁 ) 𝐸(𝜀2𝑚+1,𝓁 )
2

1 𝜀𝑚,1 𝜀𝑚,2 1 𝜀𝑚+1,1 𝜀𝑚+1,2

𝐴𝑝,1 = − 𝐸{ √ √ }, 𝐴𝑐,1 = − 𝐸{ √ √ },
2 2
𝐸(𝜀2𝑚,1 ) 𝐸(𝜀2𝑚,2 ) 𝐸(𝜀2𝑚+1,1 ) 𝐸(𝜀2𝑚+1,2 )
𝒁 𝑚,1 𝜀𝑚,2 𝒁 𝑚+1,1 𝜀𝑚+1,2
𝑨𝑝,2 = −𝐸{ √ √ }, 𝑨𝑐,2 = −𝐸{ √ √ },
𝐸(𝜀𝑚,1 ) 𝐸(𝜀𝑚,2 )
2 2 𝐸(𝜀𝑚+1,1 ) 𝐸(𝜀2𝑚+1,2 )
2

𝜀𝑚,1 𝒁 𝑚,2 𝜀𝑚+1,1 𝒁 𝑚+1,2

𝑨𝑝,3 = −𝐸{ √ √ }, 𝑨𝑐,3 = −𝐸{ √ √ }.
𝐸(𝜀𝑚,1 ) 𝐸(𝜀𝑚,2 )
2 2 𝐸(𝜀𝑚+1,1 ) 𝐸(𝜀2𝑚+1,2 )
2

Remark 1. Note that terms with 𝑨𝑝,2 and 𝑨𝑝,3 in Theorem 1 are due to the inference for conditional means in the pre-crisis period,
while terms with 𝑨𝑐,2 and 𝑨𝑐,3 in Theorem 1 are due to the inference for conditional means in the crisis period.

To develop a contagion test based on 𝜌̂𝜀 , we have to estimate 𝜎0 in the above theorem. Because neither {𝑊𝑠,𝑝 } nor {𝑊𝑠,𝑐 } is
a martingale difference sequence due to the first term in the above expressions, it is challenging to estimate 𝜎0 directly. Note that

4
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

𝑚
{𝑊𝑠,𝑝 − 𝑊𝑠,𝑐 }𝑠=10
is a strict 𝛼 -mixing sequence, we employ the overlapping block method in Shi (2000) for 𝛼 -mixing sequence to
estimate 𝜎02 . Alternatively, we can use the nonoverlapping block method in Carlstein (1986) to estimate 𝜎02 or apply the moving
√
block bootstrap method in Künsch (1989) to compute critical values for 𝑚0 (𝜌̂𝜀 − 𝜌𝜀 ). If we change C3) from 𝛼 -mixing sequences to
𝜌-mixing sequences, the employed overlapping block method for estimating the variance is still valid; see Peligrad and Shao (1995).
First, we estimate the unknown quantities of

𝐴𝑝,1 , 𝐴𝑐,1 , 𝑨𝑝,2 , 𝑨𝑐,2 , 𝑨𝑝,3 , 𝑨𝑐,3 , Γ𝑝,𝓁 , and Γ𝑐,𝓁 for 𝓁 = 1, 2
in Theorem 1 above by
𝑚0 𝑚0
1 ∑ 𝜀̂𝑚−𝑚0 +𝑡,1 𝜀̂𝑚−𝑚0 +𝑡,2 ̂ 1 ∑ 𝜀̂𝑚+𝑡,1 𝜀̂𝑚+𝑡,2
𝐴̂𝑝,1 = − , 𝐴𝑐,1 = − ,
2𝑚0 𝑡=1 𝜎
̂𝑝,1 𝜎
̂𝑝,2 2𝑚0 𝑡=1 𝜎
̂𝑐,1 𝜎 ̂𝑐,2
𝑚0 𝑚0
∑ 𝒁 𝑚−𝑚0 +𝑡,1 𝜀̂𝑚−𝑚0 +𝑡,2 ∑ 𝒁 𝑚+𝑡,1 𝜀̂𝑚+𝑡,2
̂ 𝑝,2 = − 1
𝑨 ̂ 𝑐,2 = − 1
, 𝑨 ,
𝑚0 𝑡=1 𝜎
̂𝑝,1 𝜎
̂𝑝,2 2𝑚0 𝑡=1 𝜎
̂𝑐,1 𝜎 ̂𝑐,2
𝑚0 𝑚0
∑ 𝜀̂𝑚−𝑚0 +𝑡,1 𝒁 𝑚−𝑚0 +𝑡,2 ∑ 𝜀̂𝑚+𝑡,1 𝒁 𝑚+𝑡,2
̂ 𝑝,3 = − 1
𝑨 ̂ 𝑐,3 = − 1
, 𝑨 ,
𝑚0 𝑡=1 𝜎
̂𝑝,1 𝜎
̂𝑝,2 𝑚0 𝑡=1 𝜎̂𝑐,1 𝜎 ̂𝑐,2
𝑚0 ⊤ 𝑚0 ⊤
∑ 𝒁 𝑚−𝑚0 +𝑡,𝓁 𝒁 𝑚−𝑚0 +𝑡,𝓁 ∑ 𝒁 𝑚+𝑡,𝓁 𝒁 𝑚+𝑡,𝓁
̂𝑝,𝓁 = 1
Γ ̂𝑐,𝓁 = 1
, and Γ , respectively.
𝑚0 𝑡=1 𝜎
̂𝑝,𝓁 𝜎
̂𝑝,𝓁 𝑚0 𝑡=1 𝜎
̂𝑐,𝓁 𝜎
̂𝑐,𝓁
Next, we estimate 𝑊𝑡,𝑝 and 𝑊𝑡,𝑐 for 𝑡 = 1, ⋯ , 𝑚0 by

̂𝑡,𝑝
𝜀̂𝑚−𝑚 +𝑡,1 𝜀̂ 𝑚−𝑚 +𝑡,2 1 ∑𝑚 0 𝜀̂𝑚−𝑚 +𝑠,1 𝜀̂𝑚−𝑚 +𝑠,2
𝑊 ={ 𝜎
0
̂𝑝,1 𝜎
0
̂𝑝,2
− 𝑚0 𝑠=1 𝜎
0
̂𝑝,1 𝜎
0
̂𝑝,2
}
𝜀̂𝑚−𝑚 +𝑡,1
2 𝜀̂2𝑚−𝑚 +𝑡,2
+{ 0
− 1}𝐴̂𝑝,1 + { 0
− 1}𝐴̂𝑝,1
𝜎
̂𝑝,1
2 𝜎
̂𝑝,2
2
⊤ ⊤
𝜀̂𝑚−𝑚 +𝑡,1 𝒁 𝑚−𝑚 +𝑡,1 𝜀̂𝑚−𝑚 +𝑡,2 𝒁 𝑚−𝑚 +𝑡,2
+ 0 0 ̂−1 𝑨
Γ ̂ + 0 0 ̂−1 𝑨
Γ ̂
𝜎
̂𝑝,1 𝜎
̂𝑝,1 𝑝,1 𝑝,2 𝜎
̂𝑝,2 𝜎
̂𝑝,2 𝑝,2 𝑝,3

and

̂𝑡,𝑐 𝜀̂𝑚+𝑡,1 𝜀̂ 𝑚+𝑡,2 1 ∑𝑚 0 𝜀̂𝑚+𝑠,1 𝜀̂𝑚+𝑠,2

𝑊 ={ 𝜎
̂𝑐,1 𝜎 ̂𝑐,2
− 𝑚0 𝑠=1 𝜎 ̂𝑐,1 𝜎
̂𝑐,2
}
𝜀̂2𝑚+𝑡,1 𝜀̂2𝑚+𝑡,2
+{ − 1}𝐴̂𝑐,1 + { 2 − 1}𝐴̂𝑐,1
𝜎
̂𝑐,1
2 𝜎
̂ 𝑐,2
⊤ ⊤
𝜀̂𝑚+𝑡,1 𝒁 𝑚+𝑡,1 −1 𝜀̂𝑚+𝑡,2 𝒁 𝑚+𝑡,2 −1
+ ̂ 𝑨
Γ ̂ + ̂ 𝑨
Γ ̂ .
𝜎
̂𝑐,1 𝜎
̂𝑐,1 𝑐,1 𝑐,2 𝜎
̂𝑐,2 𝜎
̂𝑐,2 𝑐,2 𝑐,3

Finally, we estimate 𝜎02 by

𝑚0 −𝑛 𝑗+𝑛
1 ∑ 1 ∑
𝜎̂ 02 = ̂ −𝑊
{𝑊 ̂𝑡,𝑐 }2 ,
𝑚0 − 𝑛 + 1 𝑗=0 𝑛 𝑡=𝑗+1 𝑡,𝑝

where 𝑛 is the window size.

Theorem 2. Assume conditions C1)-C4) hold for model (2) and 𝑛 → ∞ and 𝑛∕𝑚0 → 0 as 𝑇 → ∞. Then, under the null hypothesis of (3)
√ 𝑑
or (4), 𝑚0 𝜌̂𝜀 ∕𝜎̂ 0 → 𝑁(0, 1) as 𝑇 → ∞.

Use Theorem 2 above, we reject the no contagion null hypothesis in (3) at level 𝑎 whenever 𝑚0 𝜌̂2𝜀 ∕𝜎̂ 02 > 𝜒1,1−𝑎
2 , where 𝜒1,1−𝑎
2 is
the (1 − 𝑎)-quantile of chi-squared distribution with one degree of freedom.

Remark 2. Note that we model assets in (2) individually rather than jointly, but we incorporate dependent residuals to capture
cross-sectional dependence, following the methodology employed in copula time series models. Consequently, our measurement of
contagion naturally relies on residuals. However, when assets are modeled using a multivariate time series approach, it becomes
necessary to jointly model volatilities. In such cases, the definition of contagion should exclude interactions from both conditional
means and volatilities. Consequently, conducting a contagion test without estimating volatility presents a significant challenge.

3. Simulation study

This section examines the finite sample performance of the proposed contagion test in terms of size and power.

5
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 1
2∕3
Empirical sizes for testing (3) with 𝜌1 = ⋯ = 𝜌𝑇 = 0.1. We use the block size 𝑛 = 𝑐𝑚0
with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1 and FR2 to denote the
parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and 𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1258 0.1264 0.1291 0.0971 0.0923
normal 0.1180 0.1198 0.1209 0.0817 0.0864
200 t 0.1215 0.1225 0.1253 0.0908 0.0968
normal 0.1152 0.1168 0.1186 0.0775 0.0877
500 t 0.1018 0.1021 0.1019 0.0881 0.0884
normal 0.0958 0.0959 0.0964 0.0792 0.0837
Student
100 t 0.1461 0.1479 0.1521 0.2121 0.2148
normal 0.1267 0.1275 0.1299 0.1492 0.1629
200 t 0.1296 0.1302 0.1324 0.2294 0.2303
normal 0.1158 0.1166 0.1182 0.1559 0.1650
500 t 0.1170 0.1187 0.1200 0.2564 0.2534
normal 0.1094 0.1096 0.1098 0.1742 0.1713
Clayton
100 t 0.1313 0.1330 0.1359 0.1054 0.1004
normal 0.1232 0.1254 0.1276 0.0849 0.0913
200 t 0.1159 0.1167 0.1174 0.1032 0.0978
normal 0.1115 0.1116 0.1108 0.0878 0.0860
500 t 0.1065 0.1075 0.1083 0.1002 0.1055
normal 0.1020 0.1022 0.1015 0.0872 0.0928

Consider two AR(1)-GARCH(1,1) processes {𝑋𝑡,1 } and {𝑋𝑡,2 } in model (2) with

𝜇1 = 0, 𝜙1,1 = 0.1, 𝜆1 = 0.1, 𝜎𝑡,1

2
= 0.01 + 0.1𝑒2𝑡−1,1 + 0.8𝜎𝑡−1,1
2
,

𝜇2 = 0.01, 𝜙1,2 = 0.2, 𝜆2 = 0.2, 𝜎𝑡,2

2
= 0.02 + 0.1𝑒2𝑡−1,2 + 0.85𝜎𝑡−1,2
2
,

and {(𝜂𝑡,1 , 𝜂𝑡,2 )⊤ }𝑛𝑡=1 in (2) being independent random vectors from three bivariate copula families with the first marginal distribution
being the standardized 𝑡(5) and the second marginal distribution being the standardized 𝑡(8) (i.e., variance one) or both marginals
being 𝑁(0, 1). The employed three copula functions are:

• Gaussian copula. 𝐶(𝑢, 𝑣; 𝜌) = Φ𝜌 (Φ−1 (𝑢), Φ−1 (𝑣)), where Φ𝜌 (⋅, ⋅) is the bivariate normal distribution with means zero, variances
one, and correlation coefficient 𝜌, and Φ is the univariate standard normal distribution function.
• Student-𝑡 copula. 𝐶(𝑢, 𝑣; 𝜌) = t𝜈,𝜌 (𝑡−1
𝜈 (𝑢), 𝑡𝜈 (𝑣)), where t𝜈,𝜌 (⋅, ⋅) is the distribution function of bivariate Student-𝑡 distribution with
−1

means zero, correlation coefficient 𝜌, and 𝜈 degrees of freedom, and 𝑡𝜈 (⋅) is the univariate Student-𝑡 distribution with 𝜈 degrees
of freedom.
• Clayton copula. 𝐶(𝑢, 𝑣; 𝜌) = (𝑢−𝜌 + 𝑣−𝜌 − 1)−1∕𝜌 for 𝜌 > 0.

We generate 10000 random samples from each of the above settings with 𝑚 = 1000 and 𝑚0 ∶= 𝑇 − 𝑚 = 100, or 200, or 500. As our
method is based on paired data between pre-crisis and crisis, the effective sample size is the minimum of 𝑚 and 𝑚0 . So, we simply fix
2∕3
𝑚 = 1000 instead of changing it. We use the block size 𝑛 = 𝑐𝑚0 with 𝑐 = 0.2, 0.5, 1 for the overlapping block method in estimating
the asymptotic variance. For comparison, we implement the parametric FR test in Forbes and Rigobon (2002) using residuals from
model (2) and considering 𝑚 = 1000 and 𝑚 = 𝑚0 .
Tables 1–4 report the empirical sizes for testing the null hypothesis of (3) at level 10%, and Tables 5–8 report the empirical
powers of the proposed test for the null hypothesis of (3) at level 10%. We observe that i) the GARCH errors with normal marginals
have a more accurate coverage probability than those with 𝑡 marginals for the proposed test, especially for 𝑚0 = 100; ii) the proposed
test is more accurate as 𝑚0 increases; iii) the size of the parametric FR test in Forbes and Rigobon (2002) is severely distorted,
especially for the Student-𝑡 copula; iv) the proposed test has good power, the powers of the proposed test in Tables 6 and 8 are larger
than those in Tables 5 and 7, respectively, as the alternative hypothesis is farther away from the null hypothesis, and the comparison
with the parametric FR test in Forbes and Rigobon (2002) does not make sense as the FR test has a severely distorted size; v) the
choice of 𝑐 in using the overlapping block method for the proposed test has little impact on the test size and power; vi) powers
of the proposed test in Tables 5 and 7 do not change much when 𝑛 is increased from 100 to 500 for the Student-𝑡 copula and the
Clayton copula, which is hard to explain without deriving the Edgeworth expansions because the power depends on the departure of
alternative hypothesis to the null hypothesis, the sample size, the underlying joint distributions, and other characteristics.

6
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 2
2∕3
Empirical sizes for testing (3) with 𝜌1 = ⋯ = 𝜌𝑇 = −0.1. We use the block size 𝑛 = 𝑐𝑚0
with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1 and FR2 to denote the
parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and 𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1233 0.1233 0.1268 0.0974 0.0927
normal 0.1173 0.1182 0.1192 0.0858 0.0867
200 t 0.1227 0.1244 0.1253 0.0898 0.0955
normal 0.1167 0.1176 0.1188 0.0821 0.0869
500 t 0.1017 0.1022 0.1022 0.0887 0.0952
normal 0.0978 0.0987 0.0986 0.0788 0.0863
Student
100 t 0.1448 0.1462 0.1492 0.2183 0.2146
normal 0.1262 0.1253 0.1283 0.1524 0.1604
200 t 0.1281 0.1299 0.1313 0.2303 0.2334
normal 0.1155 0.1158 0.1179 0.1573 0.1678
500 t 0.1174 0.1178 0.1191 0.2522 0.2515
normal 0.1109 0.1113 0.1115 0.1706 0.1709
Clayton
100 t 0.1309 0.1331 0.1361 0.0899 0.0840
normal 0.1224 0.1231 0.1250 0.0791 0.0823
200 t 0.1137 0.1144 0.1160 0.0876 0.0836
normal 0.1086 0.1093 0.1098 0.0782 0.0806
500 t 0.1029 0.1026 0.1030 0.0801 0.0859
normal 0.0991 0.0994 0.0999 0.0749 0.0809

Table 3
2∕3
Empirical sizes for testing (3) with 𝜌1 = ⋯ = 𝜌𝑇 = 0.5. We use the block size 𝑛 = 𝑐𝑚0
with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1 and FR2 to denote the
parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and 𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1281 0.1282 0.1307 0.0804 0.0774
normal 0.1155 0.1176 0.1203 0.0372 0.0464
200 t 0.1248 0.1264 0.1281 0.1108 0.1195
normal 0.1148 0.1160 0.1174 0.0552 0.0551
500 t 0.1227 0.1224 0.1239 0.1932 0.1895
normal 0.0995 0.1002 0.1006 0.0974 0.0890
Student
100 t 0.1398 0.1417 0.1462 0.1422 0.1544
normal 0.1224 0.1240 0.1255 0.0943 0.1050
200 t 0.1261 0.1283 0.1281 0.1850 0.1924
normal 0.1165 0.1181 0.1190 0.1141 0.1210
500 t 0.1191 0.1190 0.1201 0.2589 0.2466
normal 0.1090 0.1094 0.1086 0.1710 0.1563
Clayton
100 t 0.1396 0.1420 0.1445 0.1204 0.1130
normal 0.1245 0.1263 0.1276 0.0819 0.0870
200 t 0.1256 0.1263 0.1293 0.1427 0.1332
normal 0.1148 0.1149 0.1153 0.0948 0.0880
500 t 0.1181 0.1168 0.1177 0.1846 0.1786
normal 0.1054 0.1057 0.1051 0.1160 0.1124

4. Empirical applications

In this section, we investigate three applications of our contagion testing method. Firstly, we examine the 2019 COVID crisis. The
dataset includes daily equity market and real estate market indices of Australia (AUS), China (CHI), Germany (GER), Japan (JAP),
the UK, and the US, covering the period from January 1, 2018, to January 31, 2022. The pre-crisis period is defined as January 1,
2018, to December 30, 2019, while the crisis period spans from December 31, 2019, to January 31, 2022. The onset of the COVID
crisis is marked on December 31, 2019, which corresponds to the first documented case of COVID-19 in China, see Fry-McKibbin et
al. (2022) for further details.
Next, we turn our attention to the second application, which investigates the 2007 subprime mortgage crisis in the US. The dataset
comprises equity market and real estate market indices of Australia (AUS), China (CHI), Hong Kong (HK), Japan (JAP), Singapore
(SIN), and the US, covering the period from January 1, 2007, to December 31, 2007. The pre-crisis period spans from January 1,

7
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 4
2∕3
Empirical sizes for testing (3) with 𝜌1 = ⋯ = 𝜌𝑇 = −0.5. We use the block size 𝑛 = 𝑐𝑚0
with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1 and FR2 to denote the
parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and 𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1264 0.1276 0.1289 0.0866 0.0819
normal 0.1140 0.1143 0.1154 0.0399 0.0446
200 t 0.1345 0.1356 0.1374 0.1272 0.1202
normal 0.1189 0.1182 0.1181 0.0573 0.0573
500 t 0.1243 0.1260 0.1267 0.1967 0.0978
normal 0.1003 0.1009 0.1007 0.1965 0.0958
Student
100 t 0.1366 0.1382 0.1428 0.1461 0.1495
normal 0.1197 0.1215 0.1243 0.0908 0.1003
200 t 0.1250 0.1270 0.1272 0.1895 0.1967
normal 0.1106 0.1116 0.1132 0.1183 0.1235
500 t 0.1202 0.1210 0.1217 0.2535 0.2528
normal 0.1104 0.1113 0.1125 0.1645 0.1528
Clayton
100 t 0.1341 0.1355 0.1370 0.1214 0.1129
normal 0.1223 0.1243 0.1257 0.0602 0.0667
200 t 0.1424 0.1244 0.1263 0.1613 0.1545
normal 0.1132 0.1135 0.1135 0.0841 0.0803
500 t 0.1178 0.1187 0.1197 0.2261 0.2238
normal 0.1005 0.1008 0.1003 0.1250 0.1195

Table 5
Empirical powers for testing (3) with 𝜌1 = ⋯ = 𝜌𝑚 = 0.1 and 𝜌𝑚+1 = ⋯ = 𝜌𝑇 = 0.05. We use
2∕3
the block size 𝑛 = 𝑐𝑚0 with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1
and FR2 to denote the parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and
𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1460 0.1470 0.1493 0.1336 0.1120
normal 0.1384 0.1390 0.1404 0.1260 0.1073
200 t 0.1595 0.1592 0.1598 0.1541 0.1329
normal 0.1569 0.1567 0.1566 0.1572 0.1289
500 t 0.2106 0.2026 0.2028 0.2287 0.1944
normal 0.2054 0.2053 0.2053 0.2397 0.2026
Student
100 t 0.1585 0.1611 0.1646 0.2462 0.2315
normal 0.1420 0.1431 0.1450 0.1844 0.1815
200 t 0.1510 0.1517 0.1524 0.2826 0.2620
normal 0.1389 0.1396 0.1398 0.2255 0.2035
500 t 0.1657 0.1669 0.1679 0.3617 0.3260
normal 0.1722 0.1729 0.1730 0.3057 0.2685
Clayton
100 t 0.1438 0.1464 0.1472 0.1216 0.1083
normal 0.1375 0.1389 0.1410 0.1096 0.1025
200 t 0.1392 0.1400 0.1427 0.1376 0.1238
normal 0.1343 0.1342 0.1358 0.1302 0.1162
500 t 0.1644 0.1630 0.1633 0.1843 0.1648
normal 0.1611 0.1608 0.1606 0.1782 0.1567

2007, to July 25, 2007, while the crisis period extends from July 26, 2007, to December 31, 2007. The selection of the pre-crisis and
crisis period cut-off dates is discussed in Forbes and Rigobon (2002) and Bond et al. (2006).
Lastly, we consider the 1997 Asian financial crisis, which has also been investigated by Forbes and Rigobon (2002). The dataset
includes daily equity market and real estate market indices of China (CHI), Hong Kong (HK), Japan (JAP), and the US, covering the
period from January 1, 1996, to June 30, 1998. The pre-crisis period encompasses January 1, 1996, to October 16, 1997, while the
crisis period spans from October 17, 1997, to June 30, 1998.3

3
All the real estate and equity indices are denoted in US dollars and collected from Bloomberg. The mnemonics of those indices are: Australia - the MXAU0RL
index and the S&P/ASX 300 index; China - the MXCN0RL index and the Shanghai stock exchange composite index; Germany - MXDE0RL index and the Deutsche
Boerse AG German stock index; Hong Kong - the MXHK0RL index and the Hang Seng index; Japan - the MXJP0RL index and the Nikkei 225 index; Singapore - the

8
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 6
Empirical powers for testing (3) with 𝜌1 = ⋯ = 𝜌𝑚 = 0.1 and 𝜌𝑚+1 = ⋯ = 𝜌𝑇 = −0.05. We
2∕3
use the block size 𝑛 = 𝑐𝑚0 with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use
FR1 and FR2 to denote the parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000
and 𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.3116 0.3120 0.3137 0.3723 0.2397
normal 0.3052 0.3078 0.3038 0.3855 0.2455
200 t 0.4408 0.4434 0.4436 0.5519 0.3838
normal 0.4467 0.4473 0.4487 0.5772 0.3940
500 t 0.7385 0.7389 0.7391 0.8189 0.6952
normal 0.7548 0.7554 0.7541 0.8412 0.7210
Student
100 t 0.2566 0.2585 0.2596 0.4280 0.3317
normal 0.2526 0.2541 0.2568 0.4014 0.2976
200 t 0.3157 0.3187 0.3187 0.5557 0.4412
normal 0.3401 0.3411 0.3415 0.5548 0.4201
500 t 0.4911 0.4904 0.4911 0.7427 0.6623
normal 0.5800 0.5806 0.5804 0.7741 0.6844
Clayton
100 t 0.2438 0.2445 0.2458 0.2794 0.1950
normal 0.2363 0.2350 0.2371 0.2743 0.1918
200 t 0.3367 0.3394 0.3406 0.4168 0.2921
normal 0.3338 0.3341 0.3360 0.4776 0.2868
500 t 0.5728 0.5728 0.5728 0.6602 0.5430
normal 0.5827 0.5824 0.5819 0.6703 0.5465

Table 7
Empirical powers for testing (3) with 𝜌1 = ⋯ = 𝜌𝑚 = 0.1 and 𝜌𝑚+1 = ⋯ = 𝜌𝑇 = 0.15. We use
2∕3
the block size 𝑛 = 𝑐𝑚0 with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1
and FR2 to denote the parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and
𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.1426 0.1445 0.1452 0.1280 0.1030
normal 0.1344 0.1356 0.1353 0.1042 0.0937
200 t 0.1581 0.1583 0.1602 0.1335 0.1188
normal 0.1541 0.1544 0.1552 0.1147 0.1063
500 t 0.1959 0.1970 0.1963 0.1580 0.1471
normal 0.1987 0.1992 0.1996 0.1455 0.1327
Student
100 t 0.1580 0.1581 0.1609 0.2371 0.2208
normal 0.1426 0.1433 0.1443 0.1658 0.1654
200 t 0.1549 0.1544 0.1563 0.2577 0.2427
normal 0.1459 0.1466 0.1472 0.1843 0.1779
500 t 0.1664 0.1665 0.1664 0.3007 0.2820
normal 0.1699 0.1704 0.1707 0.2208 0.2082
Clayton
100 t 0.1395 0.1409 0.1420 0.1235 0.1052
normal 0.1312 0.1325 0.1334 0.1002 0.0950
200 t 0.1360 0.1365 0.1385 0.1285 0.1179
normal 0.1288 0.1298 0.1306 0.1079 0.0986
500 t 0.1601 0.1598 0.1599 0.1395 0.1356
normal 0.1554 0.1553 0.1558 0.1196 0.1199

4.1. Descriptive statistics

The study converts the market prices of real estate and equity indices into percent daily log returns by calculating the difference
in the natural logarithm of each daily price and multiplying it by 100. Descriptive statistics are then calculated for both the pre-crisis
and crisis periods. The results demonstrate a significant impact of the crisis on the returns of all markets. Figs. 1–3 illustrate the
market prices of equity and real estate data, indicating a noticeable decline in index prices during the crisis period across most
countries. During such crises, it is expected that volatility will be elevated for an extended duration. Fig. 4 provides a plot of the

MXSG0RL index and the Straits Times index; the UK - the MXGB0RL index and the FTSE 100 index; and the US - the MXUS0RL index and the Dow Jones industrial
average index.

9
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 8
Empirical powers for testing (3) with 𝜌1 = ⋯ = 𝜌𝑚 = 0.1 and 𝜌𝑚+1 = ⋯ = 𝜌𝑇 = 0.25. We use
2∕3
the block size 𝑛 = 𝑐𝑚0 with 𝑐 = 0.2, 0.5, 1 for the overlapping block method. We use FR1
and FR2 to denote the parametric FR test in Forbes and Rigobon (2002) with 𝑚 = 1000 and
𝑚 = 𝑚0 , respectively.

Copula 𝑚0 Marginals 𝑐 = 0.2 𝑐 = 0.5 𝑐=1 FR1 FR2

Gaussian
100 t 0.3099 0.3113 0.3148 0.3511 0.2051
normal 0.3062 0.3074 0.3096 0.3264 0.1955
200 t 0.4469 0.4500 0.4520 0.5010 0.3248
normal 0.4535 0.4543 0.4550 0.4939 0.3256
500 t 0.7440 0.7441 0.7439 0.7113 0.5870
normal 0.7654 0.7660 0.7656 0.7468 0.6189
Student
100 t 0.2566 0.2581 0.2602 0.4089 0.2938
normal 0.2542 0.2534 0.2558 0.3556 0.2542
200 t 0.3205 0.3223 0.3229 0.5228 0.3829
normal 0.3461 0.3462 0.3473 0.4950 0.3512
500 t 0.4930 0.4941 0.4950 0.6663 0.5870
normal 0.5889 0.5883 0.5890 0.6895 0.5912
Clayton
100 t 0.2099 0.2109 0.2128 0.2339 0.1595
normal 0.1987 0.2004 0.2031 0.2009 0.1426
200 t 0.2758 0.2780 0.2778 0.3138 0.2210
normal 0.2733 0.2742 0.2739 0.2848 0.2065
500 t 0.4579 0.4592 0.4604 0.4623 0.3778
normal 0.4675 0.4685 0.4690 0.4530 0.3670

percent daily log returns during the studied data period, showing increased volatility around the mean returns after the common
shock, particularly for markets such as AUS, GER, US, JAP, UK, US, SIN, and HK. However, the returns for CHI do not consistently
exhibit a significant change in variation before and during the crisis, suggesting a lesser impact of the crisis on the CHI market
compared to markets like the US or UK, where mean returns and standard deviations undergo larger changes. Table 9 presents
descriptive statistics of the percent daily log returns for each market type and country, categorized into pre-crisis and crisis periods,
further supporting these conclusions.
On average, the mean percent daily log returns for all markets hover around 0% throughout the data period, indicating that the
net effect of the crisis is close to zero. However, the percent mean returns experienced a significant decline from the pre-crisis to
the crisis periods. These returns have larger absolute values during the crisis compared to the pre-crisis period, except for the CHI
market. Pre-crisis mean log returns are mostly positive, while mean returns during the crisis periods are predominantly negative. For
example, in the 2019 crisis data, real estate markets observed a decline in mean returns, with negative returns observed in AUS and
CHI. Conversely, equity markets show an increase in mean returns, particularly in CHI and the US.
Table 9 also reveals that volatility is generally higher during the crisis period compared to the pre-crisis period for most markets,
indicating increased data variability during crises. For instance, the mean standard deviation (SD) of real estate markets in HK and
JAP for the 1997 crisis data increases from 1.53 and 1.58, respectively, to 4.20 and 3.13 during the crisis. This rise in variability can
be attributed to heightened uncertainty and risk associated with the crisis period, leading to significant market fluctuations. Notably,
the standard deviation for 2019 JAP equity during the crisis period is relatively low compared to other markets, suggesting that this
market may be less affected by the crisis. Additionally, the crisis data exhibit greater skewness than the pre-crisis data, with larger
skewness coefficients for most markets during the crisis (mostly negative for 2019 and 2017 crisis data, and mostly positive for 1997
crisis data). This indicates a higher likelihood of extreme returns during times of crisis. These findings align with expectations, as
market volatility tends to escalate during crises, making markets more prone to extreme fluctuations.

4.2. Contagion test results

In this section, we utilize the proposed robust contagion test under the null hypothesis (4) to evaluate the existence of contagion
in real estate and equity markets during the pre-crisis and crisis periods, using the dataset introduced in the previous section.
Our robust contagion tests are conducted using an AR(p) model with unspecified heteroscedastic errors and 𝑝 = 2, 5, 10. We also
2∕3
employ the block size 𝑛 = 𝑚0 in the overlapping block method for estimating the asymptotic variance. We report results for 𝑝 = 5 in
4
Tables 10–12. Furthermore, we also perform the conventional parametric FR test (Forbes and Rigobon, 2002) to enable comparative
analysis between the results obtained from the two methods.

4
We also conduct the estimations with 𝑝 = 2 and 𝑝 = 10 which show similar results as 𝑝 = 5.

10
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Fig. 1. Plots of market prices for the real estate and equity indices for Australia, China, Germany, Japan, the UK, and the US from 1/1/2018 to 1/31/2022. The
pre-crisis period is from 1/1/2018 to 12/30/2019, and the crisis period is from 12/31/2019 to 1/31/2022.

Fig. 2. Plots of market prices for the real estate and equity indices for Australia, China, Hong Kong, Japan, Singapore, and the US from 1/1/2007 to 12/31/2007. The
pre-crisis period is from 1/1/2007 to 7/25/2007, and the crisis period is from 7/26/2007 to 12/31/2007.

4.2.1. Contagion based on our proposed robust test

Table 10 presents the results of the contagion tests during the 2019 COVID pandemic for both the real estate and equity markets.
The 𝑝-values are reported for each pair of countries. The rows corresponding to the 0-day shift show the results using the pre-crisis
period (January 1, 2018, to December 30, 2019) and the crisis period (December 31, 2019, to January 31, 2022) discussed previously.
In the real estate markets, only three out of 15 pairs of countries (AUS/GER, CHI/JAP, and GER/UK) exhibit no contagion. However,
in the equity markets, seven pairs of countries (AUS/CHI, AUS/JAP, CHI/GER, CHI/JAP, CHI/UK, CHI/US, and JAP/US) show no
contagion. Interestingly, the contagion effect is more prominent in the real estate market compared to the equity market during the
COVID-19 pandemic.
Additionally, various tests with shifts in the pre-crisis and crisis periods are conducted. These tests examine the effect of sample
size by shifting the periods by different time intervals (+10, -10, +20, and -20 days). For example, the second row in Table 10
displays the 𝑝-values when the pre-crisis period is shifted 10 days forward (+10-day). This corresponds to a change of pre-crisis
sample size from 𝑚 = 521 to 𝑚 = 531, and the crisis sample size from 𝑇 − 𝑚 = 545 to 𝑇 − 𝑚 = 535. Similarly, for the shift of -10-day
in the third row, the pre-crisis sample size is changed from 𝑚 = 521 to 𝑚 = 511, and the crisis sample size from 𝑇 − 𝑚 = 545 to
𝑇 − 𝑚 = 555. The results indicate that the contagion test is robust with small time shifting around the original outbreak of the COVID
pandemic, and most of the results align with the 0-day results.

11
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Fig. 3. Plots of market prices for the real estate and equity indices for China, Hong Kong, Japan, and the US from January 1, 1996, to June 30, 1998. The pre-crisis
period is from January 1, 1996, to October 16, 1997, and the crisis period is from October 17, 1997, to June 30, 1998.

Table 11 displays the results for the 2007 subprime mortgage crisis data in both the real estate and equity markets. In the
real estate market, the contagion pairs are AUS/CHI, AUS/SIN, CHI/JAP, CHI/SIN, and SIN/US out of the 15 pairs. For the equity
markets, the detected contagion pairs are AUS/CHI, CHI/HK, CHI/JAP, and JAP/SIN. It is noteworthy that the subprime mortgage
crisis contagion is not as widespread as expected, and the US, the origin of the crisis, does not show significant contagion to other
countries except for the SIN/US pair in the real estate market. The results for the 2007 subprime mortgage crisis are less robust with
respect to time shifting compared to the 2019 COVID pandemic.
Lastly, Table 12 presents the contagion test results for the 1997 Asian crisis. In the real estate markets, four pairs of countries
(CHI/US, HK/US, HK/US, JAP/US) out of six pairs exhibit contagion. In the equity markets, the contagion pairs are CHI/JAP, CHI/US,
and JAP/US. Unlike the 2007 crisis, the 1997 Asian crisis shows transmission from the Asian market to the US market in terms of
contagion. Similar to the 2007 crisis, the results are less robust with respect to time shifting in some pairs of countries.

4.2.2. Contagion based on the conventional FR test

In this section, the conventional FR test is conducted on the same dataset. Tables 13–15 present the empirical results of the FR
contagion test during the 2019 COVID pandemic, the 2007 subprime mortgage crisis, and the 1997 Asian crisis, respectively. As
discussed in the introduction, it is necessary to predetermine the country in which the crisis originates. We conduct the FR test
by considering each country as the crisis origin and calculating the unconditional correlation. Each result in the table represents a
source-recipient countries pair (𝑖, 𝑗). Similar to the tests in the previous section, the effect of sample size is also considered in that
the pre-crisis periods are allowed to be shifted by +10, -10, +20, and -20 days.
Table 13 reports the FR contagion effects during the 2019 COVID pandemic crisis. Compared to our proposed test, the FR test
shows some differences in the results. In the real estate markets, the FR test detects contagion from GER to AUS, while our test
does not. Additionally, the FR test shows insignificant contagion in the ordered pairs (GER, US), (UK, CHI), (US, GER), and (US, UK)
that our test considers significant. Discrepancies between the two tests are also observed in the equity market. The FR test detects
contagion in certain pairs that our test does not, and vice versa. Typically, the FR test detects contagion in the ordered pairs (CHI,
US), (GER, CHI), (JAP, US), and (UK, CHI); and shows no contagion in (GER, AUS), (GER, UK), (GER, US), (JAP, UK), (UK, AUS),
(UK, GER), (UK, JAP), (UK, US), (US, GER), and (US, JAP). These results are all opposite to ours.
Moreover, the FR test does not appear to be robust when adjusting for sample size in the study of the 2019 COVID crisis. Shifting
the sample period leads to inconsistent results. For example, the results show significant evidence of contagion from AUS to JAP
real estate markets in the case of no-shift (𝑝-value is 0.042), but no evidence of contagion is found when the crisis sample period is
shifted. Also, in the ordered pair (AUS, CHI), the 0-day case shows an insignificant 𝑝-values of 0.815. But all the shifted cases are
significant at the 10% level.
Table 14 shows limited evidence of contagion in the real estate markets during the 2007 subprime mortgage crisis using the FR
test, except for the (US, SIN). In contrast, our proposed test identifies contagion in five pairs out of 15. In the equity market, there
are contrasting results between the two tests. Our proposed test shows contagion in AUS/CHI, CHI/HK, and CHI/JAP, whereas no
contagion is found in any source-recipient directions in these countries using the FR test. The FR test also finds contagion in the
ordered pairs (AUS, JAP), (CHI, US), (JAP, AUS), and (SIN, CHI), but no contagion finds in our test among these countries. Finally,
for the 1997 Asian crisis study in Table 15, the two tests differ in the contagion results in the country pairs CHI/HK, CHI/JAP,
HK/JAP, HK/US in the real estate market. In the equity market, our test shows contagion between US and JAP, but the FR test shows
no contagion from JAP to the US.

12
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Fig. 4. Plots of the daily percent log returns of real estate and equity prices for all datasets. The top panel is the plot for the 2019 COVID dataset, the middle panel is
for the 2007 crisis, and the bottom plot is the plot for the 1997 crisis.

13
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 9
Descriptive Statistics for the Percent Log Returns of Market Crisis Data.
2019 CRISIS DATA

Full Data Pre-crisis Crisis

Market 𝑇 −1 Mean SD Min Max Skew 𝑚−1 Mean SD Min Max Skew 𝑇 −𝑚 Mean SD Min Max Skew

Estate: AUS 1,065 0.01 1.59 -18.58 9.99 -2.52 520 0.02 0.91 -4.05 2.51 -0.28 545 -0.01 2.04 -18.58 9.99 -2.29
Estate: CHI 1,065 -0.04 1.98 -8.35 7.79 0.09 520 0.02 1.96 -8.35 6.68 -0.04 545 -0.11 2.01 -8.07 7.79 0.20
Estate: GER 1,065 0.01 1.31 -8.60 7.43 -0.43 520 0.02 1.13 -5.37 5.96 -0.09 545 0.01 1.46 -8.60 7.43 -0.56
Estate: JAP 1,065 -0.01 1.29 -9.95 10.62 -0.17 520 0.00 0.88 -4.22 3.00 -0.17 545 -0.02 1.58 -9.95 10.62 -0.14
Estate: UK 1,065 0.01 1.37 -10.74 9.57 -0.28 520 0.01 1.07 -4.66 5.72 0.13 545 0.02 1.61 -10.74 9.57 -0.38
Estate: US 1,065 0.03 1.50 -18.79 8.42 -2.01 520 0.03 0.89 -3.81 3.21 -0.58 545 0.03 1.91 -18.79 8.42 -1.86

Equity: AUS 1,065 0.00 1.34 -10.81 6.54 -1.45 520 0.00 0.83 -3.71 2.66 -0.38 545 0.01 1.70 -10.81 6.54 -1.37
Equity: CHI 1,065 0.00 1.21 -9.69 6.22 -0.68 520 -0.03 1.26 -6.19 5.82 -0.35 545 0.04 1.16 -9.69 6.22 -1.06
Equity: GER 1,065 0.01 1.36 -13.94 10.47 -0.96 520 -0.01 0.96 -3.36 3.38 -0.32 545 0.03 1.65 -13.94 10.47 -1.01
Equity: JAP 1,065 0.01 1.21 -8.40 8.01 -0.11 520 0.01 0.98 -4.28 4.12 -0.52 545 0.01 1.39 -8.40 8.01 0.03
Equity: UK 1,065 0.00 1.28 -13.31 10.09 -1.18 520 -0.01 0.85 -3.36 3.00 -0.28 545 0.00 1.59 -13.31 10.09 -1.17
Equity: US 1,065 0.03 1.38 -13.84 10.76 -1.08 520 0.03 0.96 -4.71 4.86 -0.64 545 0.04 1.68 -13.84 10.76 -1.05

2007 CRISIS DATA

Full Data Pre-crisis Crisis

Market 𝑇 −1 Mean SD Min Max Skew 𝑚−1 Mean SD Min Max Skew 𝑇 −𝑚 Mean SD Min Max Skew

Estate: AUS 260 -0.01 1.75 -12.61 4.78 -1.81 147 0.05 1.25 -4.30 4.78 0.06 113 -0.10 2.24 -12.61 4.29 -1.90
Estate: CHI 260 0.38 3.05 -9.89 7.79 -0.50 147 0.67 3.27 -9.89 7.79 -0.79 113 0.00 2.72 -5.88 6.45 -0.05
Estate: HK 260 0.19 2.22 -6.28 7.20 0.07 147 0.13 1.35 -3.65 4.00 -0.05 113 0.27 3.00 -6.28 7.20 0.01
Estate: JAP 260 -0.07 2.02 -7.76 7.38 0.02 147 0.00 1.60 -5.72 4.39 -0.20 113 -0.16 2.46 -7.76 7.38 0.16
Estate: SIN 260 0.01 1.46 -5.18 3.86 -0.29 147 0.10 1.31 -5.15 3.47 -0.40 113 -0.09 1.63 -5.18 3.86 -0.15
Estate: US 260 -0.09 1.62 -5.79 4.09 -0.23 147 -0.09 1.26 -3.94 3.37 -0.14 113 -0.10 2.00 -5.79 4.09 -0.23

Equity: AUS 260 0.08 1.50 -6.15 5.09 -0.89 147 0.15 1.01 -3.82 2.42 -0.74 113 -0.01 1.96 -6.15 5.09 -0.68
Equity: CHI 260 0.28 2.16 -9.11 5.32 -1.02 147 0.35 2.32 -9.11 4.70 -1.33 113 0.20 1.95 -4.78 5.32 -0.34
Equity: HK 260 0.13 1.62 -5.20 5.73 -1.02 147 0.10 1.07 -4.09 2.66 -0.46 113 0.16 2.13 -5.20 5.73 -0.11
Equity: JAP 260 -0.02 1.13 -7.49 3.30 -1.08 147 0.02 0.86 -2.91 1.91 -0.67 113 -0.08 1.41 -7.49 3.30 -1.01
Equity: SIN 260 0.09 1.48 -4.51 6.05 -0.32 147 0.15 1.17 -4.02 2.36 -0.80 113 0.02 1.81 -4.51 6.05 -0.05
Equity: US 260 0.02 0.90 -3.35 2.52 -0.62 147 0.07 0.66 -3.35 2.07 -1.15 113 -0.03 1.14 -2.88 2.52 -0.32

1997 CRISIS DATA

Full Data Pre-crisis Crisis

Market 𝑇 −1 Mean SD Min Max Skew 𝑚−1 Mean SD Min Max Skew 𝑇 −𝑚 Mean SD Min Max Skew

Estate: CHI 651 0.10 2.52 -19.76 12.89 -0.46 468 0.12 2.79 -19.76 12.89 -0.54 183 0.03 1.64 -5.43 8.07 0.66
Estate: HK 651 -0.09 2.59 -14.21 20.69 0.41 468 0.05 1.53 -8.34 5.76 -0.51 183 -0.47 4.20 -14.21 20.69 0.63
Estate: JAP 651 -0.08 2.13 -8.96 10.63 0.26 468 0.00 1.58 -7.31 6.65 0.08 183 -0.28 3.13 -8.96 10.63 0.41
Estate: US 651 0.04 0.54 -4.23 1.78 -0.99 468 0.08 0.50 -2.50 1.78 -0.43 183 -0.06 0.64 -4.23 1.56 -1.49

Equity: CHI 651 0.14 2.22 -10.44 9.58 0.54 468 0.16 2.52 -10.44 9.58 -0.52 183 0.09 1.19 -5.11 4.88 -0.36
Equity: HK 651 -0.03 2.10 -14.76 17.32 0.26 468 0.06 1.27 -7.59 6.90 -0.68 183 -0.25 3.39 -14.76 17.32 0.48
Equity: JAP 651 -0.08 1.63 -6.46 7.68 0.14 468 -0.06 1.26 -5.00 4.38 -0.12 183 -0.14 2.33 -6.46 7.68 0.27
Equity: US 651 0.09 0.96 -7.45 4.60 -0.77 468 0.09 0.87 -3.16 3.32 -0.30 183 0.07 1.16 -7.45 4.60 -1.21

*Conversion from prices to returns reduces the sizes of total (T) and pre-crisis (m) samples by 1.

The results in this section demonstrate that the proposed robust test yields different conclusions on contagion compared to the
conventional FR test. This disparity primarily stems from the distinct approaches in estimating the asymptotic standard errors for the
difference 𝜌̂𝑝,𝜀 − 𝜌̂𝑐,𝜀 employed by the two tests. As discussed in the introduction, our method incorporates both the filtering effect and
heteroscedasticity in the asymptotic standard error calculation, leading to contrasting results when compared to the parametric FR
test. Moreover, the FR test sometimes produces unrealistic results when adjusting for time shifts around the original crisis outbreak
dates. The requirement of knowing the country where the crisis originated for the FR test can lead to unclear or confused results.
For instance, in the 2007 subprime crisis, the FR test shows only one contagion emanating from the US to SIN in the real estate
market among other pairs exhibiting contagion. Similarly, in the 1997 Asian crisis, the FR test indicated numerous contagion effects
originating from the US in both the real estate and equity markets, while failing to detect contagion from HK.

5. Conclusion

The existing literature on contagion testing assumes that data after filtering conditional means are independent and follow a
parametric distribution family, which ignores the impact of filtering conditional means. As heteroscedasticity is a stylized fact of
daily financial returns, this paper allows heteroscedastic errors with a jump from the pre-crisis to the crisis in modeling returns
and develops a contagion test without inferring heteroscedasticity and fitting a parametric distribution family, which results in a
complicated asymptotic variance of the test statistic based on testing for a change in the correlation coefficient between pre-crisis
and crisis. Hence, we employ the overlapping block method in the literature for estimating the asymptotic variance of a mixing
sequence, which requires pairing data in pre-crisis and crisis. The developed contagion test is robust against heteroscedasticity but
reduces the effective sample size. A simulation study shows that heavy tails have an impact on the test size when the effective sample
size is small. Future research to improve the size accuracy and power is to model and infer heteroscedasticities explicitly. We apply
the proposed contagion test to three crises: the 2019 COVID pandemic, the 2007 subprime mortgage crisis, and the 1997 Asian crisis.
The empirical results reveal that the proposed test yields different conclusions on contagion compared to the conventional Forbes
and Rigobon (FR) test, which is parametric and ignores the filtering effect. The results demonstrate the usefulness of our robust test
in detecting contagion.

14
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 10
P-values of contagion testing using the proposed test in the 2019 COVID crisis.

Real Estate Equity

Pre-crisis Pre-crisis
Country shift∗ CHI GER JAP UK US Country shift∗ CHI GER JAP UK US

AUS 0-day 0.000 0.131 0.042 0.000 0.001 AUS 0-day 0.815 0.000 0.874 0.000 0.000
+10-day 0.000 0.103 0.023 0.000 0.000 +10-day 0.709 0.000 0.987 0.000 0.000
-10-day 0.000 0.150 0.033 0.000 0.003 -10-day 0.894 0.000 0.883 0.000 0.000
+20-day 0.000 0.072 0.024 0.000 0.000 +20-day 0.530 0.000 0.960 0.000 0.000
-20-day 0.000 0.139 0.041 0.000 0.003 -20-day 0.895 0.000 0.820 0.000 0.000

CHI 0-day 0.000 0.669 0.008 0.003 CHI 0-day 0.398 0.616 0.860 0.230
+10-day 0.000 0.799 0.015 0.003 +10-day 0.338 0.566 0.944 0.206
-10-day 0.000 0.798 0.007 0.006 -10-day 0.383 0.446 0.877 0.262
+20-day 0.000 0.833 0.007 0.000 +20-day 0.609 0.697 0.666 0.086
-20-day 0.000 0.680 0.005 0.006 -20-day 0.388 0.546 0.854 0.272

GER 0-day 0.091 0.161 0.058 GER 0-day 0.001 0.001 0.042
+10-day 0.066 0.122 0.054 +10-day 0.001 0.000 0.021
-10-day 0.122 0.191 0.069 -10-day 0.001 0.002 0.033
+20-day 0.092 0.095 0.034 +20-day 0.001 0.000 0.022
-20-day 0.139 0.328 0.071 -20-day 0.004 0.005 0.026

JAP 0-day 0.023 0.026 JAP 0-day 0.015 0.130

+10-day 0.018 0.008 +10-day 0.017 0.098
-10-day 0.025 0.032 -10-day 0.023 0.165
+20-day 0.013 0.005 +20-day 0.017 0.062
-20-day 0.023 0.035 -20-day 0.025 0.171

UK 0-day 0.000 UK 0-day 0.000

+10-day 0.000 +10-day 0.000
-10-day 0.000 -10-day 0.000
+20-day 0.000 +20-day 0.000
-20-day 0.000 -20-day 0.000

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

Table 11
P-values of contagion testing using the proposed test in the 2007 subprime mortgage crisis.

Real Estate Equity

Pre-crisis Pre-crisis
Country shift∗ CHI HK JAP SIN US Country shift∗ CHI HK JAP SIN US

AUS 0-day 0.001 0.343 0.839 0.045 0.429 AUS 0-day 0.006 0.113 0.236 0.257 0.582
+10-day 0.004 0.377 0.372 0.007 0.907 +10-day 0.015 0.233 0.589 0.642 0.708
-10-day 0.009 0.212 0.855 0.159 0.437 -10-day 0.033 0.107 0.367 0.202 0.529
+20-day 0.109 0.579 0.272 0.383 0.986 +20-day 0.254 0.493 0.020 0.863 0.971
-20-day 0.001 0.255 0.635 0.198 0.582 -20-day 0.073 0.057 0.397 0.116 0.317

CHI 0-day 0.170 0.005 0.027 0.685 CHI 0-day 0.019 0.017 0.286 0.358
+10-day 0.314 0.019 0.065 0.295 +10-day 0.021 0.044 0.463 0.972
-10-day 0.334 0.033 0.133 0.596 -10-day 0.126 0.051 0.518 0.098
+20-day 0.760 0.764 0.266 0.326 +20-day 0.117 0.363 0.665 0.762
-20-day 0.124 0.023 0.120 0.283 -20-day 0.149 0.185 0.740 0.172

HK 0-day 0.975 0.976 0.793 HK 0-day 0.199 0.863 0.557

+10-day 0.201 0.974 0.434 +10-day 0.293 0.305 0.520
-10-day 0.808 0.997 0.984 -10-day 0.770 0.620 0.359
+20-day 0.706 0.132 0.591 +20-day 0.170 0.082 0.840
-20-day 0.812 0.687 0.549 -20-day 0.462 0.626 0.610

JAP 0-day 0.488 0.829 JAP 0-day 0.083 0.414

+10-day 0.128 0.834 +10-day 0.988 0.731
-10-day 0.509 0.844 -10-day 0.315 0.511
+20-day 0.285 0.876 +20-day 0.023 0.175
-20-day 0.486 0.996 -20-day 0.249 0.309

SIN 0-day 0.058 SIN 0-day 0.953

+10-day 0.902 +10-day 0.155
-10-day 0.108 -10-day 0.605
+20-day 0.432 +20-day 0.580
-20-day 0.170 -20-day 0.847

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

15
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 12
P-values of contagion testing using the proposed test in the 1997 Asian crisis.

Real Estate Equity

Pre-crisis Pre-crisis
Country shift∗ HK JAP US Country shift∗ HK JAP US

CHI 0-day 0.172 0.243 0.003 CHI 0-day 0.290 0.044 0.042
+10-day 0.307 0.285 0.957 +10-day 0.060 0.036 0.384
-10-day 0.140 0.234 0.003 -10-day 0.454 0.103 0.081
+20-day 0.497 0.146 0.802 +20-day 0.082 0.075 0.545
-20-day 0.157 0.246 0.007 -20-day 0.727 0.156 0.123

HK 0-day 0.002 0.006 HK 0-day 0.143 0.205

+10-day 0.007 0.482 +10-day 0.506 0.001
-10-day 0.001 0.023 -10-day 0.037 0.362
+20-day 0.021 0.942 +20-day 0.923 0.042
-20-day 0.004 0.019 -20-day 0.044 0.439

JAP 0-day 0.008 JAP 0-day 0.012

+10-day 0.013 +10-day 0.043
-10-day 0.014 -10-day 0.044
+20-day 0.247 +20-day 0.238
-20-day 0.001 -20-day 0.024

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

Table 13
P-values of contagion testing using the Forbes and Rigobon (2002) test in the 2019 COVID crisis.

Real Estate Equity

Pre-crisis Recipient (𝑗 ) Recipient (𝑗 )

Source (𝑖) shift∗ AUS CHI GER JAP UK US AUS CHI GER JAP UK US

AUS 0-day n.a. 0.000 0.131 0.042 0.000 0.001 n.a. 0.815 0.000 0.874 0.000 0.000
+10-day n.a. 0.012 0.696 0.247 0.371 0.027 n.a. 0.055 0.913 0.020 0.865 0.308
-10-day n.a. 0.011 0.647 0.187 0.455 0.036 n.a. 0.058 0.833 0.020 0.821 0.299
+20-day n.a. 0.015 0.746 0.245 0.413 0.036 n.a. 0.056 0.903 0.018 0.898 0.361
-20-day n.a. 0.011 0.596 0.289 0.678 0.041 n.a. 0.054 0.932 0.015 1.000 0.292

CHI 0-day 0.000 n.a. 0.001 0.153 0.022 0.000 0.176 n.a. 0.617 0.352 0.501 0.028
+10-day 0.000 n.a. 0.001 0.164 0.022 0.000 0.177 n.a. 0.580 0.464 0.512 0.036
-10-day 0.000 n.a. 0.000 0.158 0.023 0.000 0.169 n.a. 0.578 0.334 0.547 0.030
+20-day 0.015 n.a. 0.746 0.245 0.413 0.036 0.056 n.a. 0.903 0.018 0.898 0.361
-20-day 0.011 n.a. 0.596 0.289 0.678 0.041 0.054 n.a. 0.932 0.015 1.000 0.292

GER 0-day 0.081 0.004 n.a. 0.013 0.659 0.244 0.354 0.009 n.a. 0.049 0.436 0.953
+10-day 0.100 0.005 n.a. 0.013 0.692 0.272 0.369 0.008 n.a. 0.061 0.460 0.862
-10-day 0.075 0.003 n.a. 0.012 0.673 0.205 0.307 0.008 n.a. 0.044 0.482 0.999
+20-day 0.119 0.004 n.a. 0.015 0.791 0.322 0.332 0.008 n.a. 0.101 0.415 0.837
-20-day 0.058 0.002 n.a. 0.010 0.935 0.205 0.338 0.009 n.a. 0.048 0.391 0.976

JAP 0-day 0.051 0.908 0.091 n.a. 0.033 0.009 0.238 0.240 0.016 n.a. 0.102 0.047
+10-day 0.058 0.887 0.089 n.a. 0.035 0.009 0.257 0.191 0.018 n.a. 0.098 0.049
-10-day 0.037 0.890 0.087 n.a. 0.034 0.008 0.247 0.258 0.014 n.a. 0.099 0.038
+20-day 0.058 0.878 0.094 n.a. 0.037 0.007 0.250 0.185 0.036 n.a. 0.167 0.073
-20-day 0.058 0.844 0.077 n.a. 0.040 0.008 0.201 0.301 0.017 n.a. 0.094 0.046

UK 0-day 0.026 0.249 0.696 0.008 n.a. 0.004 0.628 0.041 0.223 0.369 n.a. 0.566
+10-day 0.021 0.256 0.682 0.009 n.a. 0.003 0.625 0.038 0.227 0.932 n.a. 0.622
-10-day 0.032 0.272 0.628 0.009 n.a. 0.005 0.555 0.036 0.254 0.364 n.a. 0.548
+20-day 0.024 0.260 0.627 0.008 n.a. 0.003 0.611 0.040 0.198 0.543 n.a. 0.686
-20-day 0.085 0.268 0.339 0.016 n.a. 0.008 0.765 0.040 0.158 0.372 n.a. 0.656

US 0-day 0.012 0.055 0.612 0.026 0.124 n.a. 0.020 0.826 0.609 0.053 0.099 n.a.
+10-day 0.010 0.040 0.597 0.024 0.102 n.a. 0.019 0.896 0.698 0.061 0.111 n.a.
-10-day 0.013 0.049 0.659 0.023 0.126 n.a. 0.017 0.826 0.566 0.042 0.092 n.a.
+20-day 0.013 0.028 0.542 0.019 0.114 n.a. 0.024 0.879 0.751 0.092 0.143 n.a.
-20-day 0.012 0.048 0.669 0.022 0.123 n.a. 0.013 0.791 0.547 0.047 0.098 n.a.

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

16
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Table 14
P-values of contagion testing using the Forbes and Rigobon (2002) test in the 2007 subprime mortgage crisis.

Real Estate Equity

Pre-crisis Recipient (𝑗 ) Recipient (𝑗 )

Source (𝑖) shift∗ AUS CHI HK JAP SIN US AUS CHI HK JAP SIN US

AUS 0-day n.a. 0.661 0.722 0.362 0.775 0.291 n.a. 0.484 0.571 0.052 0.403 0.284
+10-day n.a. 0.383 0.964 0.420 0.888 0.061 n.a. 0.367 0.568 0.059 0.526 0.575
-10-day n.a. 0.477 0.497 0.397 0.961 0.150 n.a. 0.614 0.787 0.033 0.279 0.213
+20-day n.a. 0.421 0.813 0.540 0.626 0.173 n.a. 0.317 0.841 0.019 0.486 0.941
-20-day n.a. 0.572 0.562 0.438 0.928 0.169 n.a. 0.271 0.707 0.040 0.204 0.178

CHI 0-day 0.371 n.a. 0.632 0.952 0.511 0.345 0.845 n.a. 0.778 0.885 0.201 0.083
+10-day 0.119 n.a. 0.817 0.289 0.184 0.889 0.530 n.a. 0.754 0.433 0.179 0.494
-10-day 0.236 n.a. 0.720 0.625 0.372 0.137 0.959 n.a. 0.597 0.788 0.314 0.114
+20-day 0.170 n.a. 0.625 0.330 0.182 0.738 0.363 n.a. 0.718 0.898 0.124 0.807
-20-day 0.385 n.a. 0.854 0.538 0.798 0.131 0.406 n.a. 0.919 0.372 0.218 0.233

HK 0-day 0.784 0.393 n.a. 0.683 0.412 0.666 0.400 0.149 n.a. 0.176 0.233 0.710
+10-day 0.544 0.697 n.a. 0.586 0.128 0.994 0.259 0.184 n.a. 0.236 0.102 0.980
-10-day 0.958 0.454 n.a. 0.748 0.267 0.971 0.583 0.243 n.a. 0.143 0.334 0.836
+20-day 0.416 0.887 n.a. 0.840 0.078 0.972 0.208 0.261 n.a. 0.725 0.125 0.490
-20-day 0.832 0.552 n.a. 0.570 0.123 0.997 0.535 0.224 n.a. 0.097 0.348 0.892

JAP 0-day 0.469 0.961 0.926 n.a. 0.142 0.397 0.064 0.739 0.278 n.a. 0.076 0.975
+10-day 0.532 0.445 0.794 n.a. 0.151 0.480 0.051 0.333 0.319 n.a. 0.120 0.816
-10-day 0.524 0.694 0.957 n.a. 0.119 0.387 0.042 0.541 0.236 n.a. 0.082 0.736
+20-day 0.421 0.472 0.928 n.a. 0.976 0.533 0.020 0.731 0.251 n.a. 0.047 0.555
-20-day 0.587 0.591 0.855 n.a. 0.125 0.427 0.053 0.296 0.166 n.a. 0.103 0.792

SIN 0-day 0.333 0.645 0.639 0.152 n.a. 0.102 0.389 0.073 0.387 0.062 n.a. 0.730
+10-day 0.318 0.325 0.831 0.182 n.a. 0.120 0.507 0.093 0.312 0.142 n.a. 0.787
-10-day 0.578 0.496 0.840 0.120 n.a. 0.048 0.269 0.110 0.521 0.066 n.a. 0.649
+20-day 0.633 0.270 0.724 0.690 n.a. 0.562 0.603 0.091 0.539 0.066 n.a. 0.964
-20-day 0.644 0.843 0.808 0.120 n.a. 0.033 0.177 0.095 0.551 0.084 n.a. 0.759

US 0-day 0.297 0.550 0.477 0.377 0.064 n.a. 0.320 0.102 0.766 0.980 0.672 n.a.
+10-day 0.072 0.879 0.757 0.467 0.080 n.a. 0.622 0.403 0.835 0.801 0.584 n.a.
-10-day 0.149 0.291 0.752 0.365 0.030 n.a. 0.306 0.111 0.826 0.815 0.515 n.a.
+20-day 0.186 0.777 0.739 0.522 0.335 n.a. 0.618 0.620 0.619 0.529 0.732 n.a.
-20-day 0.156 0.242 0.792 0.393 0.019 n.a. 0.255 0.187 0.877 0.871 0.596 n.a.

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

Table 15
P-values of contagion testing using the Forbes and Rigobon (2002) test in the 1997 Asian Crisis.

Real Estate Equity

Pre-crisis Recipient (𝑗 ) Recipient (𝑗 )

Source (𝑖) shift∗ CHI HK JAP US CHI HK JAP US

CHI 0-day n.a. 0.028 0.040 0.000 n.a. 0.136 0.000 0.000
+10-day n.a. 0.264 0.085 0.006 n.a. 0.009 0.000 0.245
-10-day n.a. 0.038 0.055 0.000 n.a. 0.192 0.000 0.000
+20-day n.a. 0.203 0.019 0.002 n.a. 0.004 0.000 0.181
-20-day n.a. 0.140 0.165 0.000 n.a. 0.721 0.022 0.000

HK 0-day 0.311 n.a. 0.770 0.762 0.320 n.a. 0.538 0.758

+10-day 0.615 n.a. 0.663 0.714 0.214 n.a. 0.685 0.012
-10-day 0.326 n.a. 0.606 0.740 0.332 n.a. 0.494 0.836
+20-day 0.572 n.a. 0.725 0.885 0.181 n.a. 0.965 0.069
-20-day 0.452 n.a. 0.569 0.719 0.525 n.a. 0.378 0.916

JAP 0-day 0.427 0.710 n.a. 0.067 0.120 0.993 n.a. 0.105
+10-day 0.578 0.546 n.a. 0.113 0.136 0.744 n.a. 0.122
-10-day 0.458 0.868 n.a. 0.078 0.190 0.974 n.a. 0.176
+20-day 0.396 0.583 n.a. 0.268 0.151 0.892 n.a. 0.307
-20-day 0.592 0.895 n.a. 0.060 0.379 0.812 n.a. 0.210

US 0-day 0.000 0.118 0.011 n.a. 0.061 0.195 0.047 n.a.

+10-day 0.121 0.043 0.002 n.a. 0.858 0.000 0.013 n.a.
-10-day 0.000 0.111 0.014 n.a. 0.027 0.227 0.091 n.a.
+20-day 0.065 0.194 0.016 n.a. 0.734 0.000 0.065 n.a.
-20-day 0.000 0.105 0.011 n.a. 0.019 0.261 0.112 n.a.

*Crisis period is shifted in exact opposite direction -(+) to pre-crisis period +(-).

17
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Acknowledgements

We thank an associate editor and two reviewers for their helpful comments, which improved the presentation of this paper.

Appendix A. Proofs of Theorems

√
̃ 𝑡,𝓁 = 𝒁 𝑡,𝓁 ∕
Proof of Theorem 1. Define 𝒁 𝐸(𝜀2𝑡,𝓁 ). Using the weak law of large numbers, the central limit theory for 𝛼 -mixing
sequences, and the Taylor expansion, we have

⎧ 1 ∑𝑚 ̃ ̃⊤
𝑝
⎪ Γ𝑝,𝓁,𝑇 ∶= 𝑚0 𝑡=𝑚−𝑚0 +1 𝒁 𝑡,𝓁 𝒁 𝑡,𝓁 → Γ𝑝,𝓁 ,
⎪̂ −1 1 ∑𝑚 𝜀𝑡,𝓁
̃ −1∕2
⎪ 𝜻 𝑝,𝓁 − 𝜻 𝓁 = Γ𝑝 𝑚0 𝑡=𝑚−𝑚0 +1 √𝐸(𝜀2 ) 𝒁 𝑡,𝓁 + 𝑜𝑝 (𝑚0 ),
⎪ 𝑡,𝓁
⎪Γ 1 ∑𝑇 𝑝
̃ ⊤ → Γ𝑐,𝓁 ,
̃ 𝑡,𝓁 𝒁
𝑐,𝓁,𝑇 ∶= 𝑚 𝑡=𝑚+1 𝒁 𝑡,𝓁
⎨ (5)
1 ∑𝑇
0
𝜀𝑡,𝓁
⎪̂𝜻 𝑐,𝓁 − 𝜻 𝓁 = Γ−1 √ ̃ 𝑡,𝓁 + 𝑜𝑝 (𝑚−1∕2 ),
𝒁
𝑐 𝑚0 𝑡=𝑚+1
⎪ 𝐸(𝜀2𝑡,𝓁 ) 0
⎪ ̂ ⊤
⎪ 𝜀̂𝑡,𝓁 = 𝜀𝑡,𝓁 − (𝜻 𝑝,𝓁 − 𝜻 𝓁 ) 𝒁 𝑡,𝓁 for 𝑡 = 𝑚 − 𝑚0 + 1, ⋯ , 𝑚,
⎪ 𝜀̂𝑡,𝓁 = 𝜀𝑡,𝓁 − (̂
𝜻 𝑐,𝓁 − 𝜻 𝓁 )⊤ 𝒁 𝑡,𝓁 for 𝑡 = 𝑚 + 1, ⋯ , 𝑇 .
⎩
Because
𝑚
∑ 𝑚
∑
1 1 −1∕2
𝜀̂2𝑡,𝓁 = 𝜀2𝑡,𝓁 + 𝑜𝑝 (𝑚0 ),
𝑚0 𝑡=𝑚−𝑚0 +1
𝑚0 𝑡=𝑚−𝑚0 +1

it follows from (5) and the Taylor expansion that

1 ∑𝑚 𝜀̂𝑡,1 𝜀̂𝑡,2
𝜌̂𝑝,𝜀 = 𝑚0 𝑡=𝑚−𝑚0 +1
√ ∑𝑚 √
−1 ∑𝑚
𝑚−10
𝜀
̂
𝑠=𝑚−𝑚0 +1 𝑠,1
2 𝑚 0 𝑠=𝑚−𝑚0 +1 𝜀 ̂2𝑠,2
1 ∑ 𝑚 𝜀̂𝑡,1 𝜀̂𝑡,2
= √ ∑𝑚 √ ∑𝑚
𝑚0 𝑡=𝑚−𝑚0 +1
𝑚−10 𝑠=𝑚−𝑚0 +1 𝜀𝑠,1
2 𝑚−1
0 𝑠=𝑚−𝑚0 +1 𝜀𝑠,2
2

1 ∑𝑚 𝜀𝑡,1 (𝜻̂𝑝,1 −𝜻 1 )⊤ 𝒁 𝑡,1 −1∕2

= 𝑚0 𝑡=𝑚−𝑚0 +1 {
√ ∑𝑚 − √ + 𝑜𝑝 (𝑚0 )}×
𝑚−1 𝜀
𝑠=𝑚−𝑚0 +1 𝑠,1
2 𝐸(𝜀 2 )
0 𝑡,1
𝜀 (̂
𝜻 −𝜻 )⊤ 𝒁
− 𝑝,2√ 2 𝑡,2 + 𝑜𝑝 (𝑚0 )}
−1∕2
{ √ ∑𝑡,2𝑚
𝑚−1 𝑠=1 𝜀2𝑠,2 𝐸(𝜀2𝑡,2 )
∑ 𝑚 𝜀𝑡,1 𝜀𝑡,2 (𝜻̂ −𝜻 )⊤ 𝒁 𝜀
= 1
𝑚0 𝑡=𝑚−𝑚0 +1 {
√ ∑𝑚 √
−1 ∑𝑚
− 𝑝,1√ 1 𝑡,1 √ 𝑡,2
𝑚−1 𝜀
𝑠=𝑚−𝑚0 +1 𝑠,1
2 𝑚 𝜀
𝑠=𝑚−𝑚0 +1 𝑠,2
2 𝐸(𝜀2 ) 𝐸(𝜀 2 )
0 0 𝑡,1 𝑡,2
𝜀𝑡,1 (𝜻̂𝑝,2 −𝜻 2 )⊤ 𝒁 𝑡,2 −1∕2
−√ √ } + 𝑜𝑝 (𝑚0 )
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
1 ∑𝑚 𝜀𝑡,1 𝜀𝑡,2
= √ √
𝑚0 𝑡=𝑚−𝑚0 +1
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
∑𝑚 𝜀𝑡,1 𝜀𝑡,2 ∑𝑚 𝜀2𝑡,1
− 12 { 𝑚1 𝑡=𝑚−𝑚0 +1
√ √ }{ 𝑚1 𝑡=𝑚−𝑚0 +1 ( 𝐸(𝜀2 ) − 1)}
0 𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 ) 0 𝑡,1
∑𝑚 𝜀𝑡,1 𝜀𝑡,2 ∑ 𝜀2𝑡,2
− 12 { 𝑚1 𝑡=𝑚−𝑚0 +1
√ √ }{ 𝑚1 𝑚 𝑡=𝑚−𝑚0 +1 ( 𝐸(𝜀2 ) − 1)}
0 𝐸(𝜀𝑡,2 )
𝐸(𝜀2𝑡,1 ) 2 0 𝑡,2
∑𝑚 ̂ ⊤𝒁̃ 𝑡,1 √ 𝜀𝑡,2
− 𝑚1 𝑡=𝑚−𝑚0 +1 (𝜻 𝑝,1 − 𝜻 1 )
0 𝐸(𝜀2𝑡,2 )
∑𝑚 𝜀𝑡,1
̂ ̃ 𝑡,2 + 𝑜𝑝 (𝑚−1∕2 ).
− 𝑚1 𝑡=𝑚−𝑚0 +1
√ (𝜻 𝑝,2 − 𝜻 2 )⊤ 𝒁 0
0 𝐸(𝜀2𝑡,1 )

By the weak law of large numbers for 𝛼 -mixing sequences, we have

⎧ 1 1 ∑𝑚 𝜀𝑡,1 𝜀𝑡,2 𝑝
⎪ 𝐴𝑝,1,𝑇 ∶= − 2 { 𝑚0 𝑡=𝑚−𝑚0 +1 √ 2 √ 2 } → 𝐴𝑝,1 ,
𝐸(𝜀𝑡,1 ) 𝐸(𝜀𝑡,2 )
⎪
⎪𝑨 1 ∑𝑚 ̃ 𝜀𝑡,2 𝑝
∶= − 𝒁 √ → 𝑨𝑝,2 ,
⎨ 𝑝,2,𝑇 𝑚0 𝑡=𝑚−𝑚0 +1 𝑡,1
𝐸(𝜀2𝑡,2 )
⎪ ∑ 𝜀𝑡,1 𝑝
⎪ 𝑨𝑝,3,𝑇 ∶= − 1 𝑚 √ ̃ 𝑡,2 → 𝑨𝑝,3 .
𝒁
𝑚0 𝑡=𝑚−𝑚0 +1
⎪ 𝐸(𝜀2𝑡,1 )
⎩
Therefore,

18
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

1 ∑𝑚 𝜀𝑡,1 𝜀𝑡,2
𝜌̂𝑝,𝜀 = 𝑚0 𝑡=𝑚−𝑚0 +1
√ √
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
∑𝑚 𝜀2𝑡,1
+𝐴𝑝,1 { 𝑚1 𝑡=𝑚−𝑚0 +1 ( 𝐸(𝜀2 ) − 1)}
0 𝑡,1 (6)
∑𝑚 𝜀2𝑡,2
+𝐴𝑝,1 { 𝑚1 𝑡=𝑚−𝑚0 +1 ( 𝐸(𝜀2 ) − 1)}
0 𝑡,2
+(𝜻̂ 𝑝,1 − 𝜻 1 )⊤ 𝑨𝑝,2 + (𝜻̂ 𝑝,2 − 𝜻 2 )⊤ 𝑨𝑝,3 + 𝑜𝑝 (𝑚0
−1∕2
).
Define
𝜀𝑡,1 𝜀𝑡,2 𝜀𝑡,1 𝜀𝑡,2
𝑊𝑡−𝑚+𝑚0 ,𝑝 = {√ √ − 𝐸( √ √ )}
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 ) 𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
𝜀2𝑡,1 𝜀2
+{ − 1}𝐴𝑝,1 + { 𝑡,22 − 1}𝐴𝑝,1
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀𝑡,2 )
𝜀
+ √ 𝑡,1 𝒁 ̃ ⊤ Γ−1 𝑨𝑝,2 + √ 𝜀𝑡,2 𝒁 ̃ ⊤ Γ−1 𝑨𝑝,3
𝑡,1 𝑝,1 𝑡,2 𝑝,2
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )

for 𝑡 = 𝑚 − 𝑚0 + 1, ⋯ , 𝑚. By (6), we have

𝑚0
√ 1 ∑
𝑚0 {̂
𝜌𝑝,𝜀 − 𝜌𝑝,𝜀 } = √ 𝑊𝑡,𝑝 + 𝑜𝑝 (1). (7)
𝑚0 𝑡=1
Similarly, using the weak law of large numbers for 𝛼 -mixing sequences, we have

⎧ 1 1 ∑𝑇 𝜀𝑡,1 𝜀𝑡,2 𝑝
⎪ 𝐴𝑐,1,𝑇 ∶= − 2 { 𝑚0 𝑡=1+𝑚 √ 2 𝐻( √ 2 )} → 𝐴𝑐,1 ,
𝐸(𝜀𝑡,1 ) 𝐸(𝜀𝑡,2 )
⎪
⎪𝑨 1 ∑𝑇 ̃ 𝜀𝑡,2 𝑝
∶= − 𝒁 √ → 𝑨 𝑐,2 ,
⎨ 𝑐,2,𝑇 𝑚0 𝑡=1+𝑚 𝑡,1
𝐸(𝜀2𝑡,2 )
⎪ ∑ 𝜀𝑡,1 𝑝
⎪ 𝑨𝑐,3,𝑇 ∶= − 1 𝑇 √ ̃ 𝑡,2 → 𝑨𝑐,3 .
𝒁
𝑚0 𝑡=1+𝑚
⎪ 𝐸(𝜀2𝑡,1 )
⎩
Define
𝜀𝑡,1 𝜀𝑡,2 𝜀𝑡,1 𝜀𝑡,2
𝑊𝑡−𝑚,𝑐 = {√ √ − 𝐸( √ √ )}
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 ) 𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
𝜀2𝑡,1 𝜀2
+{ − 1}𝐴𝑐,1 + { 𝑡,22 − 1}𝐴𝑐,1
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀𝑡,2 )
𝜀
+ √ 𝑡,1 𝒁 ̃ ⊤ Γ−1 𝑨𝑐,2 + √ 𝜀𝑡,2 𝒁 ̃ ⊤ Γ−1 𝑨𝑐,3
𝑡,1 𝑐,1 𝑡,2 𝑐,2
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )

for 𝑡 = 𝑚 + 1, ⋯ , 𝑇 . Like the proof of (6),

1 ∑𝑇 𝜀𝑡,1 𝜀𝑡,2
𝜌̂𝑐,𝜀 = 𝑚0 𝑡=1+𝑚
√ √
𝐸(𝜀2𝑡,1 ) 𝐸(𝜀2𝑡,2 )
∑𝑇 𝜀2𝑡,1
+𝐴𝑐,1 { 𝑚1 𝑡=1+𝑚 ( 𝐸(𝜀2 ) − 1)}
0 𝑡,1
∑𝑇 𝜀2𝑡,2
+𝐴𝑐,1 { 𝑚1 𝑡=1+𝑚 ( 𝐸(𝜀2 ) − 1)}
0 𝑡,2
+(𝜻̂ 𝑐,1 − 𝜻 1 )⊤ 𝑨𝑐,2 + (𝜻̂ 𝑐,2 − 𝜻 2 )⊤ 𝑨𝑐,3 + 𝑜𝑝 (𝑇 −1∕2 ),
implying that
𝑚0
√ 1 ∑
𝑚0 {̂
𝜌𝑐,𝜀 − 𝜌𝑐,𝜀 } = √ 𝑊𝑡,𝑐 + 𝑜𝑝 (1). (8)
𝑚0 𝑡=1
Hence, it follows from (7) and (8) that
√ 𝑑
𝜌𝜀 − 𝜌𝜀 ) → 𝑁(0, 𝜎02 ). □
𝑚0 (̂

Proof of Theorem 2. It follows from Theorem 1.1 and Remark 1.3 in Shi (2000). □

References

Ang, A., Chen, J., 2002. Asymmetric correlations of equity portfolios. J. Financ. Econ. 63 (3), 443–494.
Bae, K.-H., Karolyi, G.A., Stulz, R.M., 2003. A new approach to measuring financial contagion. Rev. Financ. Stud. 16 (3), 717–763.
Bekaert, G., Ehrmann, M., Fratzscher, M., Mehl, A., 2014. The global crisis and equity market contagion. J. Finance 69 (6), 2597–2649.
Bekaert, G., Hodrick, R.J., Zhang, X., 2009. International stock return comovements. J. Finance 64 (6), 2591–2626.

19
E. Aboagye, S.I.-M. Ko, C.C. Lo et al. Journal of Economic Dynamics and Control 159 (2024) 104804

Bond, S.A., Dungey, M., Fry, R., 2006. A web of shocks: crises across Asian real estate markets. J. Real Estate Finance Econ. 32, 253–274.
Bouri, E., Gupta, R., Wang, S., 2022. Nonlinear contagion between stock and real estate markets: International evidence from a local gaussian correlation approach.
Int. J. Financ. Econ. 27 (2), 2089–2109.
Boyer, B.H., Gibson, M.S., Loretan, M., 1999. Pitfalls in tests for changes in correlations. Federal Reserve Board(597), IFS Discussion Paper.
Carlstein, E., 1986. The use of subseries values for estimating the variance of a general statistic from a stationary sequence. Ann. Stat. 14 (3), 1171–1179.
Chan, J.C., Fry-McKibbin, R., Hsiao, C.Y.-L., 2019. A regime switching skew-normal model of contagion. Stud. Nonlinear Dyn. Econom. 23 (1), 20170001.
Dungey, M., Gajurel, D., 2015. Contagion and banking crisis–international evidence for 2007–2009. J. Bank. Finance 60, 271–283.
Forbes, K.J., Rigobon, R., 2002. No contagion, only interdependence: measuring stock market comovements. J. Finance 57 (5), 2223–2261.
Fry-McKibbin, R., Greenwood-Nimmo, M., Hsiao, C.Y.-L., Qi, L., 2022. Higher-order comoment contagion among G20 equity markets during the COVID-19 pandemic.
Finance Res. Lett. 45, 102150.
Fry-McKibbin, R., Hsiao, C.Y.-L., 2018. Extremal dependence tests for contagion. Econom. Rev. 37 (6), 626–649.
Fry-McKibbin, R., Hsiao, C.Y.-L., Martin, V.L., 2019. Joint tests of contagion with applications. Quant. Finance 19 (3), 473–490.
Fry-McKibbin, R., Hsiao, C.Y.-L., Martin, V.L., 2021. Measuring financial interdependence in asset markets with an application to eurozone equities. J. Bank. Fi-
nance 122, 105985.
Fry-McKibbin, R., Martin, V.L., Tang, C., 2010. A new class of tests of contagion with applications. J. Bus. Econ. Stat. 28 (3), 423–437.
Hsiao, C.Y.-L., Morley, J., 2022. Debt and financial market contagion. Empir. Econ. 62 (4), 1599–1648.
Islam, R., Volkov, V., 2022. Contagion or interdependence? Comparing spillover indices. Empir. Econ. 63 (3), 1403–1455.
Kangogo, M., Dungey, M., Volkov, V., 2023. Changing vulnerability in Asia: contagion and spillovers. Empir. Econ. 64 (5), 2315–2355.
Künsch, H.R., 1989. The jackknife and the bootstrap for general stationary observations. Ann. Stat. 17, 1217–1261.
Longin, F., Solnik, B., 2001. Extreme correlation of international equity markets. J. Finance 56 (2), 649–676.
Loretan, M., English, W.B., 2000. Evaluating correlation breakdowns during periods of market volatility. Available at SSRN 231857.
Peligrad, M., Shao, Q.M., 1995. Estimation of the variance of partial sums of 𝜌-mixing random variables. J. Multivar. Anal. 52 (1), 140–157.
Ronn, E.I., Sayrak, A., Tompaidis, S., 2009. The impact of large changes in asset prices on intra-market correlations in the domestic and international markets. Financ.
Rev. 44 (3), 405–436.
Shi, S., 2000. Estimation of the variance for strongly mixing sequences. Appl. Math. J. Chin. Univ. 15, 45–54.
Wu, B., Huang, D., Chen, M., 2023. Estimating contagion mechanism in global equity market with time-zone effect. Financ. Manag.
Xu, H., Li, S., Tian, Z., 2023. Does foreign equity investment impact the spillover effect of industries in China? N. Am. J. Econ. Finance 68, 101955.
Zhu, K., Ling, S., 2015. LADE-based inference for ARMA models with unspecified and heavy-tailed heteroscedastic noises. J. Am. Stat. Assoc. 110, 784–794.

20