Вы находитесь на странице: 1из 6

Finance Research Letters xxx (xxxx) xxxx

Contents lists available at ScienceDirect

Finance Research Letters


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

When does the stock market recover from a crisis?☆


Yanglin Li, Shaoping Wang , Qing Zhao

School of Economics, Huazhong University of Science and Technology, Wuhan, China

ARTICLE INFO ABSTRACT

Keywords: This paper develops a backward infimum augmented Dickey-Fuller (BIADF) test for identifying
Bubble-led crisis the financial crises and estimating the market recovery dates in conjunction with a backward sup
Market recovery date ADF (BSADF) test to estimate the bubble origination and termination dates. Consistency of these
Market normalcy date estimators is established. The new method is not only easy-to-implement and computa-
Double recursive regression
tionally efficient, but can accurately estimate the market recovery date. Simulations show that
when = 1, the rates of successful crisis detection are over 98.7%. We apply the tests to the US
JEL classification:
and China’s stock markets, and the results show that our method can successfully capture typical
C12
C22, financial crises. For the US stock market, the BIADF test successfully detects the Black Monday
crash of 1987 and the Dot Com crash in the early 2000s. The recovery dates are February 1991,
and September 2005, respectively. For China’s stock market, the 2008 financial crisis and the
2015 stock market disaster are detected, and the corresponding recovery dates are April 29,
2011, and June 24, 2016.

1. Introduction

The econometric identifications of financial bubbles and crises have drawn much attention since the seminal work of
Phillips et al. (2011). Related literature generally characterizes financial bubbles in terms of explosive autoregressive time series
models and detects bubbles based on recursive right-tailed unit root tests. Phillips et al. (2011) proposed a supremum augmented
Dickey-Fuller (SADF) test for detecting a single explosive bubble. Homm and Breitung (2012) provided alternative tests based on
other unit root statistics, such as the Kim- and Chow-type statistics. Phillips et al. (2015a, 2015b, PSY) extended the SADF test to the
backward sup ADF (BSADF) test for detecting multiple bubbles and estimating the origination and termination of each bubble. The
model of PSY, where the bubble implodes abruptly within one sample period, only considers the bubble expansion and ignores the
crisis led by the bubble implosion.
However, observing the trajectory of financial crises, we usually find a complex and lengthy process of market correction and
recovery. Harvey et al. (2016, HLST) modeled a bubble-led crisis as an interim stationary process to capture the mean reversion
process from bubbles collapse to market normalcy. However, hypothesis testing of HLST requires a prior known crisis regime. The
existing common methods topically identify crises based on large price drops (e.g. Malliaropulos and Migiakis, 2018; Wei et al.,
2019). But when will the stock market recover from a crisis? The answer is important, but there is little literature in the field.
Phillips and Shi (2018) addressed this concern by proposing a reverse-sample use of the BSADF test for detecting crises and estimating
the market recovery date, which is defined as the date when asset prices return to market normalcy. Nevertheless, detecting bubbles
and crises requires two double recursive regressions, i.e., the BSADF procedure and the reverse regression procedure, which is


We are grateful to the editor and the anonymous referee for their constructive comments. The authors are supported by NSFCno. 71671070.

Corresponding author.
E-mail address: wangshaoping@hust.edu.cn (S. Wang).

https://doi.org/10.1016/j.frl.2020.101642
Received 23 December 2019; Received in revised form 30 April 2020; Accepted 5 June 2020
1544-6123/ © 2020 Elsevier Inc. All rights reserved.

Please cite this article as: Yanglin Li, Shaoping Wang and Qing Zhao, Finance Research Letters,
https://doi.org/10.1016/j.frl.2020.101642
Y. Li, et al. Finance Research Letters xxx (xxxx) xxxx

complex and computationally inefficient. Our motivation is to improve the method of estimating the market recovery date and
increase the crisis detection rate.
In this study, we propose a backward infimum ADF (BIADF) test to identify the crises and estimate the market recovery date, and
use the BSADF test together to identify the bubble origination and termination dates. The consistency of these date estimators is
established, and our required conditions for consistent are less restrictive than those in Phillips and Shi (2018). We further investigate
the performance of our proposed procedure via Monte Carlo simulation. Our method is simple and easy-to-implement; meanwhile, it
yields accurate estimates of the market recovery date when detecting the crisis led by the bubble implosion. Due to its improved
performance under sudden collapses, our procedure successfully detects the Black Monday crash of 1987, while the reverse regression
procedure of Phillips and Shi (2018) did not.
The remainder of the paper is organized as follows. Section 2 presents the model and theoretical basis of the BSADF and BIADF
tests. Section 3 reports the simulation results of the proposed procedure. Section 4 gives the empirical results of the new procedure
applied to the US and China’s stock markets. Section 5 concludes.

2. A test for determining the market recovery date

Following Phillips and Shi (2018), we introduce a market recovery date to capture the return from a crisis to market normalcy. Let
the bubble origination, termination, and market recovery dates be denoted by Te, Tc, and Tr, so that B = [Te, Tc ] is the bubble episode,
C = (Tc , Tr ) is the crisis period, N0 = [1, Te ) and N1 = (Tr , T ] represent normal market periods. Thus, the data generating process is
specified as:
dT + yt 1 + t , t N0 N1,
yt = T yt 1 + t , t B
T yt 1 + t , t C (1)
2
where d is a constant, T is the sample size, η > 1/2, y0 = Op (1), εt ~ N(0, σ ), T = 1 + c1 T and T = 1 c2 T with c1, c2 > 0 and
α, β ∈ [0, 1). The speeds of bubble expansion and implosion are controlled by α and β, respectively. The larger α, the smoother the
bubble process, and the larger β, the slower the implosion rate during the crisis period C.
The PSY strategy for detecting the origination and termination of the bubble behavior is to conduct a BSADF test for each
observation t ≔ ⌊fT⌋, where f is the sample fraction corresponding to t. Let f1 and f2 be the (fractional) starting and ending points of
the ADF regression:
k
yt = µ + ( 1) yt 1 + i yt i + et
i=1 (2)
f
where t = f1 T , …, f2 T . The corresponding ADF statistic is denoted as ADF f12 . Then, the BSADF statistic is calculated by:
f
BSADFf (f0 ) = sup ADF f12 , f [f0 , 1]
f1 [0, f f0 ], f2 = f (3)
where f0 is the smallest window size in (2).
Given the double recursive left-sided test introduced in Leybourne et al. (2007) that can consistently partition the sample data
into separate unit root and stationary processes, we propose a new real-time strategy, a backward infimum augmented Dickey-Fuller
statistic,
f
BIADFf (f0 ) = inf ADF f12, f [f0 , 1]
f1 [0, f f0 ], f2 = f (4)
for detecting the presence of the crises and estimating the recovery date. We use the BIADF test together with the BSADF test to detect
financial bubbles and crises. We can see from (3) and (4) that based on one double recursive regression, the values of the BSADF and
BIADF statistics can be obtained simultaneously to determine the estimates of the bubble origination, termination, and market
recovery dates, which is more convenient than the reverse regression procedure proposed by Phillips and Shi (2018).
The market recovery date is identified by comparing the BIADF statistic with its left-tail critical value, that is, the market recovery
date is calculated as the first chronological observation whose BIADF statistic is below its corresponding critical value. The fractional
dates of bubble origination, termination, and market recovery dates are denoted by fe, fc, and fr with corresponding estimates f^e , f^c ,
and f^ , which are calculated as follows:
r

f^e = inf {f : BSADFf (f0 ) > scv T }


f [f0 ,1] (5)

f^c = inf {f : BSADFf (f0 ) < scv T }


f [f^e + LT ,1] (6)

f^r = inf {f : BIADFf (f0 ) < icv T }


f [f^c ,1] (7)
where scv T is the (1 T ) 100% critical value of the BSADFf(f0) statistic, icv T is the βT100% critical value of the BIADFf(f0) statistic,

2
Y. Li, et al. Finance Research Letters xxx (xxxx) xxxx

LT = log(T )/T is a minimum duration of bubbles, where δ is a sample-frequency dependent parameter. For f^e , f^c and f^r defined in
(5)–(7), we have the following theorem.
Theorem 1. Under the alternative hypothesis of mildly explosive behavior in model (1), if the conditions

T /2 scv T
+ 0 if < and 1 + <2
scv T T1 /2
,
1 scv T
+ 0 otherwise .
scv T T1 /2

hold, we have f^e fe and f^c


p p
fc as T → ∞. And if

T /2 |icv T |
+ 0 if > and 1 + <2
|icv T | T1/2
T (1 + )/2 |icv T |
+ 0 if > and 1 + >2
|icv T | T1/2
1 ,
T2+ |icv T |
+ 0 if < and 1 + >2
|icv T | T (1 + )/2

1 |icv T |
+ 1 0 if < and 1 + <2
|icv T | T2+

we have f^r
p
fr as T → ∞.
Remark 1. Theorem 1 shows that, the conditions for consistent estimation of fe, fc, and fr depend on the relative size of α and β.
These conditions are less restrictive than those in Phillips and Shi (2018).
Remark 2. Although the consistent conditions in Theorem 1 seem strict and complex, highly accurate estimates of fe, fc, and fr in
later simulations show that these conditions are always satisfied. So in empirical research, these date estimators can be considered to
be consistent with the true dates, which is the same as when the BSADF test is applied in the empirical study (see, Giglio et al., 2016;
Anundsen et al., 2016; Tsvetanov et al., 2016).

3. Simulation evidence

This section performs simulations to explore the finite sample performances of the BIADF and BSADF tests and to evaluate the
accuracy of date estimators. For (1), we set: y0 = 100 and d = c1 = c2 = 1. We consider two values of σ, = 1 and = 6.79, the latter is
used in Phillips and Shi (2018) and Phillips et al. (2015a). The corresponding critical values are obtained from simulations with 2000
replications and T = 0.05. The minimum window size is set according to f0 = 0.01 + 1.8/ T with T = 100 . Successful detection of a
bubble and a crisis can be defined as f < f^ < f and f < f^ < f . We restrict the duration of bubble expansion to d = (f
e e c e c r f ) T 3, BT c e
and there is no restriction on the crisis duration dCT = (fr fc ) T .
Based on the conditions in Theorem 1, we set four specifications for (α, β). Case 1: ( , ) = (0.6, 0.1), Case 2: ( , ) = (0.6, 0.4),
Case 3: ( , ) = (0.4, 0.6), Case 4: ( , ) = (0.1, 0.6) . Table 1 reports the rates of successful bubble and crisis detection, along with the
bias and standard deviation (given in parentheses) of f^ , f^ , and f^ under these four cases. Let = 0.6, f = 0.3, f = 0.5 and f = 0.7.
e c r e c r
As evident in Table 1, the BSADF and BIADF tests are robust to the specifications for (α, β), which means that the consistent
conditions in Theorem 1 are always satisfied and the specifications for (α, β) has no material impact on the bubble and crisis
detection. Additionally, the supplementary material shows that the BSADF and BIADF tests are robust to the drift in the unit root
process when η > 0.5. All of the detection rates are higher than 90%, except Case 2 with = 6.79 . The high accuracy of f^e , f^c , and f^r

Table 1
Bubble and crisis detection rates of the BSADF and BIADF tests, and estimation of fe, fc, and fr based on the full sample with different specifications
for (α, β). Figures in parentheses are standard deviations.
Case 1 Case 2 Case 3 Case 4

=1
Bubble detection rate 0.957 0.943 0.950 0.960
f^ f 0.02 (0.02) 0.02 (0.02) 0.01 (0.02) 0.01 (0.01)
e e

f^c fc 0.00 (0.02) 0.01 (0.03) 0.05 (0.03) 0.04 (0.03)


Crisis detection rate 0.993 0.987 0.989 0.991
f^ f −0.02 (0.00) −0.02 (0.00) −0.02 (0.00) −0.02 (0.00)
r r
= 6.79
Bubble detection rate 0.900 0.879 0.958 0.970
0.06 (0.04) 0.06 (0.06) 0.03 (0.03) 0.02 (0.01)
f^
e fe

f^c fc 0.00 (0.03) 0.01 (0.04) 0.05 (0.03) 0.04 (0.03)


Crisis detection rate 0.922 0.917 0.985 0.992
f^ f −0.02 (0.02) −0.02 (0.02) −0.02 (0.02) −0.02 (0.00)
r r

3
Y. Li, et al. Finance Research Letters xxx (xxxx) xxxx

is also satisfactory regardless of the relative size of α and β. In particular, there is a two-observation bias (earlier than the true date) in
estimating f , and the standard deviation of f^ is close to zero, which do not depend on the relative size of α and β.
r r
Now, we discuss the impact of the crisis duration on the BIADF test. Following Phillips and Shi (2018), we consider three collapse
patterns: sudden collapse ( = 0.1, and dCT = 0.01T ), disturbing collapse ( = 0.5, and dCT = 0.1T ), and smooth collapse ( = 0.9,
and dCT = 0.2T ). The other parameters are set to: = 0.6, = 0.6, fe = 0.3, and fc = 0.5. Table 2 displays the results of the BSADF
and BIADF tests under three collapse patterns. These results show that as σ increases from 1 to 6.79, the bubble detection rates
decrease from 95% to 85%. The BSADF test provides higher accurate estimates of fc for sudden collapses than for disturbing collapses,
while f^e is unaffected by the collapse patterns.
Table 2 shows that when = 1, the crisis detection rates are above 98% for three collapse patterns. The bias ( f^ f ) is smaller
r r
with a longer crisis duration. When = 6.79, the crisis detection rate for sudden collapses is 87.4%, which has improved significantly
compared with that (44%) in Phillips and Shi (2018). Moreover, the BIADF test provides greater accurate estimates of the market
recovery dates for smooth collapses than sudden collapses, which is contrary to the results in Phillips and Shi (2018). For example,
when = 6.79, for sudden collapses, f^ f = 0.17, and for smooth collapses, f^
r r f = 0.09. Thus, given the fact that smooth col-
r r
lapses are more common than sudden collapses, our proposed test is more practical.

4. When does the stock market recover from a crisis?

In this section, we apply the BIADF test in conjunction with the BSADF test to the US and China’s stock markets for detecting the
bubble-led crises and estimating the corresponding bubble origination date fe, bubble termination date fc and market recovery date fr.

4.1. The US stock market

Phillips et al. (2015a) identified the bubble episodes of the US stock market by using the S&P 500 stock price index, but they did
not consider the existence of the crises. Phillips and Shi (2018) used the NASDAQ composite index to detect the bubbles and crises of
the US stock market. In this study, we use the S&P 500 stock price index as a representative of the US stock market due to the
availability of data. The real S&P 500 stock price index and dividend series obtained from Shuping Shi’s homepage1 are sampled
monthly over the period from January 1973 to August 2013, including 488 observations. We date the bubble and crisis episodes of
the US stock market using the S&P 500 price-dividend ratio that is plotted in Fig. 1 by the solid (black) line.
Results are reported in Fig. 1. The left (right) panel for bubble (crisis) identification plots the BSADF (BIADF) statistic sequence
against the corresponding 95% (5%) critical value sequence, which is obtained by 2000 Monte Carlo simulation with T = 488. We set
the smallest window size to 44 observations according to f0 = 0.01 + 1.8/ T and set a lag order of one for the ADF regression. In
Fig. 1(a), there are two speculative bubbles in the US stock market in 1986M05–1987M09 and 1995M11–2001M02, which are
similar to the results in Phillips and Shi (2018). This indicates that the typical bubble and crisis episodes of the S&P500 stock price
index are the same as those of the NASDAQ composite index.
From Fig. 1(b), the BIADF test successfully detects the Black Monday crash in October 1987, which is a typical sudden collapse.
But Phillips and Shi (2018) did not detect this crash and explained it was perhaps because their crisis detection rate is lower when the
collapse is rapid with a short duration. Thus, our successful detection of this crash verifies that the proposed BIADF test can improve
the crisis detection rate. The recovery date of the Black Monday crash is February 1991. The results of the BIADF test also show that
the Dot Com crash in the early 2000s is detected, which is considered as a smooth collapse. This crash lasts 56 months and the
recovery date is September 2005, which is later than that (April 2004) in Phillips and Shi (2018) but is consistent with
Chen et al. (2018) that used the sample periods 1995–2005 to test the effect of the Dot Com crash. Besides, the identified crisis period
no longer overlaps with the identified bubble period because the repeated estimation of the bubble termination date in Phillips and
Shi (2018) is avoided.

4.2. China’s stock market

China’s stock market, as an emerging market, has attracted wide attention (e.g. He et al., 2019; Zhao et al., 2020; Deng et al.,
2017). Here we explore the bubble-led crises and the market recovery dates of China’s stock market, represented by the HS300 stock
index. We use the weekly data from January 7, 2005, to December 29, 2017, taken from the Wind Economic Database. Samples of
631 weeks in total are plotted in Fig. 2 by the black lines. Fig. 2 reports the results of the BIADF and BSADF tests. The finite sample
critical sequences are obtained by 2000 Monte Carlo simulations with 631 observations and the smallest window size is 51 ob-
servations.
The results in Fig. 2 provide evidence that there are two bubble-led crises in China’s stock market, corresponding to the financial
crisis of 2008 and the stock market disaster of 2015. Each of the bubble-led crises starts with exponential growth in prices. For the
financial crisis of 2008, the bubble originates on April 21, 2006, and collapses on January 25, 2008. There is a 22-month period of
exuberance followed by a 39-month implosion, and the recovery date is April 29, 2011. The bubble that triggered the stock market
disaster of 2015 begins on November 28, 2014, and bursts on June 19, 2015. The continuous increase lasts 7 months, but the duration

1
https://sites.google.com/site/shupingshi/home/codes

4
Y. Li, et al. Finance Research Letters xxx (xxxx) xxxx

Table 2
Bubble and crisis detection rates of the BSADF and BIADF tests, and estimation of fe, fc, and fr based on the full sample for different
collapse patterns. Figures in parentheses are standard deviations.
Sudden Disturbing Smooth

=1
Bubble detection rate 0.957 0.943 0.950
0.02(0.02) 0.02(0.02) 0.02(0.02)
f^ ef e

f^c fc 0.00(0.02) 0.02(0.03) 0.10(0.04)


Crisis detection rate 0.993 0.987 0.989
f^ f 0.17(0.00) 0.08(0.00) 0.04(0.02)
r r
= 6.79
Bubble detection rate 0.845 0.855 0.865
f^ f 0.05(0.06) 0.06(0.06) 0.06(0.06)
e e

f^c fc −0.01(0.05) 0.01(0.06) 0.08(0.08)


Crisis detection rate 0.874 0.871 0.392
f^ f 0.17(0.02) 0.08(0.03) 0.09(0.09)
r r

Fig. 1. The identified bubble and crisis episodes using the BSADF (left) and BIADF (right) tests for the S&P 500 price-dividend ratio.

Fig. 2. The identified bubble and crisis episodes using the BSADF (left) and BIADF (right) tests for the HS300 stock index.

5
Y. Li, et al. Finance Research Letters xxx (xxxx) xxxx

of the crisis episode is 12 months. The corresponding recovery date is June 24, 2016. This is the first research on the recovery from a
crisis to market normalcy for China’s stock market.

5. Conclusion

This study develops the BIADF test to identify the crises and estimate the market recovery date, and uses the BSADF test together
to detect the bubble origination and termination dates. We demonstrate the consistency of these date estimators. Simulations cor-
roborate that our method is superior to the reversed regression procedure proposed by Phillips and Shi (2018), and when = 1 the
successful crisis detection rates are over 98.7%. The proposed procedure is simpler and can improve the crisis detection rate.
Applying to the US stock market, the BIADF test successfully detects the Black Monday crash of 1987 and the Dot Com crash in the
early 2000s. The recovery dates are February 1991, and September 2005, respectively. For China’s stock market, our application
concludes that there are two typical financial crises during 2005–2017, namely, the financial crisis of 2008 and the stock market
disaster of 2015. The recovery dates are April 29, 2011, and June 24, 2016, respectively. We can learn a lesson that a short-term
boom in the stock market often comes with a subsequent long-term recession.

CRediT authorship contribution statement

Yanglin Li: Methodology, Software, Data curation, Writing - original draft, Conceptualization. Shaoping Wang: Writing - review
& editing, Supervision, Funding acquisition. Qing Zhao: Writing - review & editing, Validation.

Supplementary material

Supplementary Material. The supplement provides detailed technical proofs of Theorem 1.


Supplementary material associated with this article can be found, in the online version, at 10.1016/j.frl.2020.101642

References

Anundsen, A.K., Gerdrup, K., Hansen, F., Kragh-Sørensen, K., 2016. Bubbles and crises: the role of house prices and credit. J. Appl. Econom. 31 (7), 1291–1311.
https://doi.org/10.1002/jae.2503.
Chen, H.-C., Chou, R.K., Lu, C.-L., 2018. Saving for a rainy day: evidence from the 2000 dot-com crash and the 2008 credit crisis. J. Corp. Finance 48, 680–699. https://
doi.org/10.1016/j.jcorpfin.2017.12.025.
Deng, Y., Girardin, E., Joyeux, R., Shi, S., 2017. Did bubbles migrate from the stock to the housing market in china between 2005 and 2010? Pac. Econ. Rev. 22 (3),
276–292. https://doi.org/10.1111/1468-0106.12230.
Giglio, S., Maggiori, M., Stroebel, J., 2016. No-bubble condition: model-free tests in housing markets. Econometrica 84 (3), 1047–1091. https://doi.org/10.3982/
ECTA13447.
Harvey, D.I., Leybourne, S.J., Sollis, R., Taylor, A.R., 2016. Tests for explosive financial bubbles in the presence of non-stationary volatility. J. Empir. Finance 38,
548–574. https://doi.org/10.1016/j.jempfin.2015.09.002.
He, Q., Qian, Z., Fei, Z., Chong, T.T.-L., 2019. Do speculative bubbles migrate in the chinese stock market? Empir. Econ. 56 (2), 735–754. https://doi.org/10.1007/
s00181-017-1369-4.
Homm, U., Breitung, J., 2012. Testing for speculative bubbles in stock markets: a comparison of alternative methods. J. Financ. Econom. 10 (1), 198–231. https://doi.
org/10.1093/jjfinec/nbr009.
Leybourne, S., Kim, T.-H., Taylor, A.R., 2007. Detecting multiple changes in persistence. Stud. Nonlinear Dyn. Econom. 11 (3). https://doi.org/10.2202/1558-3708.
1370.
Malliaropulos, D., Migiakis, P., 2018. The re-pricing of sovereign risks following the global financial crisis. J. Empir. Finance 49, 39–56. https://doi.org/10.1016/j.
jempfin.2018.09.003.
Phillips, P.C., Shi, S.-P., 2018. Financial bubble implosion and reverse regression. Econ. Theory 34 (4), 705–753. https://doi.org/10.1017/S0266466617000202.
Phillips, P.C., Wu, Y., Yu, J., 2011. Explosive behavior in the 1990s Nasdaq: when did exuberance escalate asset values? Int. Econ. Rev. 52 (1), 201–226. https://doi.
org/10.1111/j.1468-2354.2010.00625.x.
Phillips, P.C.B., Shi, S., Yu, J., 2015a. Testing for multiple bubbles: historical episodes of exuberance and collapse in the S&P 500. Int. Econ. Rev. 56 (4), 1043–1078.
https://doi.org/10.1111/iere.12132.
Phillips, P.C.B., Shi, S., Yu, J., 2015b. Testing for multiple bubbles: limit theory of realtime detectors. Int. Econ. Rev. 56 (4), 1079–1134. https://doi.org/10.1111/iere.
12131.
Tsvetanov, D., Coakley, J., Kellard, N., 2016. Bubbling over! the behaviour of oil futures along the yield curve. J. Empir. Finance 38, 516–533. https://doi.org/10.
1016/j.jempfin.2015.08.009.
Wei, Y., Qin, S., Li, X., Zhu, S., Wei, G., 2019. Oil price fluctuation, stock market and macroeconomic fundamentals: evidence from china before and after the financial
crisis. Finance Res. Lett. 30, 23–29. https://doi.org/10.1016/j.frl.2019.03.028.
Zhao, Z., Wen, H., Li, K., 2020. Identifying bubbles and the contagion effect between oil and stock markets: new evidence from china. Econ. Model. https://doi.org/10.
1016/j.econmod.2020.02.018.

Вам также может понравиться