Chapter 13

Structural Breaks, Efficiency, and Volatility: An Empirical Investigation of Southeast Asian Frontier Markets

P. Andrikopoulos*
D.L.T. Anh**
M.K. Newaz*
*    Coventry Business School, Coventry, United Kingdom
**    School of Banking and Finance, National Economics University, Hanoi, Vietnam

Abstract

The purpose of this study is to examine the presence of structural breaks, efficiency, and volatility in the frontier markets of Southeast Asia. Using daily and weekly returns of four composite indices (VN Index, HNX Index, LSX Index, and CSX Index), we measure the presence of price volatility in the markets of Vietnam, Laos, and Cambodia. Our findings from all tests—unit root, autocorrelation, run, and variance ratio—suggest that the Vietnamese stock market is weak-form inefficient, while in the cases of Laos and Cambodia the empirical statistics produce inconclusive results. Furthermore, symmetric volatility models are statistically significant in both daily and weekly series, implying that the impacts of positive and negative news or shocks are the same in magnitude.

Keywords

random walk
market efficiency
Southeast Asia frontier markets

1. Introduction

This study investigates stock price behavior of the Southeast Asian (SEA) frontier markets of Vietnam, Laos, and Cambodia. To do so, we are testing stock market efficiency of these markets, and especially the extent to which stock prices follow a random walk. The contribution of this paper is therefore twofold. First, from an academic perspective, it complements prior evidence on the issue of informational efficiency for these markets using out-of-sample data, especially covering the period following the 2007–09 global financial crisis. Second, from a professional investment viewpoint, assessing the level of stock market efficiency of these countries allows better understanding of their economies and of the soundness of their financial systems, as in the presence of market efficiency stock market prices should accurately reflect companies’ actual performance. Furthermore, the understanding of how securities markets perform allows relevant governments to adopt appropriate policies to stimulate growth and capital investments, leading to the improvement of the country’s economic environment.
In brief, using alternative testing procedures, our results suggest that all three markets under examination are currently weak-form inefficient. Furthermore, the degree of informational inefficiency (or not) varies on the basis of the methodology adopted and the frequency of the examined stock return series. For example, although all markets appear to be influenced by long memory dynamics and symmetric volatility, testing for randomness in the return series indicate a marginal degree of efficiency in the case of the newer stock markets of Cambodia and Laos.
The rest of the chapter is structured as follows. Section 2 presents a review of key prior literature on market efficiency and empirical evidence from emerging and frontier markets. Data and methodology are provided in Section 3; Section 4 reports the findings of the empirical tests. Finally, conclusions are drawn in Section 5.

2. Literature Review

According to the efficient market hypothesis (Fama, 1965), stock returns in a weak-form efficient market follow a random walk model, meaning that they are unpredictable and uncorrelated to past price information. Although the concept of random walk was first mentioned in a study by Pearson (1905), it was Maurice Kendall (1953) who demonstrated these mathematical properties in a number of stock market indices. Subsequent studies (Cowles, 1960; Osborne, 1962) further confirm these findings, leading to the development of the random walk hypothesis (Samuelson, 1965) that emphasizes the impossibility of investors forecasting stocks’ rates of return in a market with fully available information. Working intensely on the concept of stock market efficiency since the early 1960s, Eugene Fama introduced this groundbreaking hypothesis in his seminal paper in 1970, by defining the term “efficient market” as “a market at which prices always “fully reflect” available information” (Fama, 1970, p. 383). Therefore, under this condition all new information should be instantly reflected in stock prices.
The concept of market efficiency has been extensively investigated in different countries under different market conditions and alternative test procedures. Nonetheless, findings are not always conclusive. For example, in a study of the US market over the period 1963–73, Sharma and Kennedy (1977) show that US equity prices follow a random walk, evidence that was further corroborated by Szakmary et al. (1999) for NASDAQ. On the other hand, by running specification tests using weekly return series for the period 1962–85, Lo and MacKinlay (1988) produce evidence of serial correlation during the whole period and within all subperiods. These findings corroborated an earlier study by Fama and French (1987), who reject the presence of random walks in the US stock market during the period 1926–40. In a subsequent study, Fama and French (1988) suggest that equity returns can be predicted from their factor components.
Outside the US market, and especially for Europe, Worthington and Higgs (2004) report the existence of randomness in the stock markets of Germany, Ireland, Portugal, Sweden, and the United Kingdom but not in the markets of Austria, Belgium, Denmark, Finland, France, Greece, Italy, Netherlands, Norway, Spain, and Switzerland. These findings are also supported by later studies (Borges, 2011; Mishra, 2012; Smith, 2012).
In a follow-up study, Worthington and Higgs (2005) confirm that the markets of Hong Kong, New Zealand, and Japan are weak-form efficient, but their findings are refuted by Suleman et al. (2010), who show that all stock prices of Hong Kong, Singapore, Japan, and Australia do not follow a random walk. Similar findings are documented for all the emerging markets of Pakistan, India, Sri Lanka, China, Indonesia, Malaysia, Philippine, Thailand, and Bangladesh (Suleman et al. 2010; Nisar and Hanif;  2012). Finally, a similar picture is reported for the emerging and frontier markets of Latin America, with all equity markets of Argentina, Brazil, Chile, Columbia, Mexico, Peru, and Venezuela being weak-form inefficient (Worthington and Higgs, 2003; Righi and Ceretta, 2011).
In contrast to other emerging and frontier markets in various regions, research on randomness in equity prices of SEA frontier markets is limited. Notably, there are no prior studies in the literature on weak-form efficiency in the Laos and Cambodia stock markets. Regarding Vietnam, there are few studies that provide consistent results. In an examination of the market index (VN Index) and five individual stocks listed on Ho Chi Minh Stock Exchange (HOSE) for the period 2000–04, Loc (2006) provides evidence of positive autocorrelation in the share price, hence rejecting the random walk hypothesis. This evidence was further corroborated by later studies using a number of testing methodologies and time periods (Guidi and Gupta, 2011; Aumeboonsuke, 2012; Vinh and Thao, 2013). Nonetheless, in a more recent study by Phan and Zhou (2014) the authors conclude that the Vietnamese stock market does not follow a random walk during the whole period, but in contrast to prior studies, there is a significant improvement in efficiency during the period 2009–13. Our study adds to this literature by examining the newly formed markets of Laos and Cambodia for the first time.

3. Data and Methodology

3.1. The Data

All three markets under examination are relatively new. Vietnam’s stock market officially came into operation on Jul. 20, 2000, with the establishment of the HOSE, while 5 years later the Hanoi Stock Exchange (HNX) was inaugurated. The indicators for stock price volatility in the HOSE and the HNX are the VN Index and the HNX Index, respectively. While the VN Index is a collection of large-cap listed companies with more than VND 120 billion of equity capital, the HNX Index consists of medium and small-cap companies with a minimum capital of VND 30 billion. Over the years, the Vietnamese stock market has seen a remarkable surge of new companies listed, mostly driven by the simple registration procedures and the small capital requirements. The number of listed companies in the market has risen from only 5 equities in 2000 to 1240 securities at the end of 2014. In contrast, the other two markets are considerably smaller in size and in magnitude. The Laos Securities Exchange (LSX) was the first capital market in the Lao PDR and officially opened in Oct. 2010. Although the purpose of opening the LSX was to attract foreign capital to fund the long-term capital requirements of local companies, 5 years after its establishment there are still only four securities listed on the market. Finally, one of the most recent and smallest markets in the SEA region is the Cambodia Securities Exchange (CSX), which was launched in Apr. 2012 with only two securities [Phnom Penh Water Supply Authority and Grand Twins International (Cambodia) Plc] that were listed at the end of 2014. For the purposes of this study, we use daily and weekly prices of the four market indices (VN, HNX, LSX, and CSX) from the period of their establishment up to the end of 2014. All price series are obtained from Thomson’s Datastream database. As shown in Table 13.1, the study analyzes a total of 3432 (VN Index) and 2194 (HNX Index) daily observations for the Vietnamese market. As regards the stock markets of Laos and Cambodia, our time series for returns is considerably smaller, consisting of 983 and 616 daily observations, respectively, covering the period from Jan. 2011 to Dec. 2014 (Laos) and Mar. 2012 to Dec. 2014 (Cambodia).

Table 13.1

Summary of Sample Size for the Index of Each SEA Frontier Stock Market

Index Frontier market Sample period No. of daily observations No. of weekly observations
VN Index Vietnam Jul. 28, 2000 to Dec. 31, 2014 3432 741
HNX Index Vietnam Jan. 4, 2006 to Dec. 31, 2014 2194 462
LSX Index Lao PDR Jan. 12, 2011 to Dec. 31, 2014 983 170
CSX Index Cambodia Apr. 19, 2012 to Dec. 31, 2014 616 141

Daily and weekly returns are estimated as

Rt=log pt−log pt−1=logptpt−1

image(13.1)
where Rt is the index return at time t, while pt and pt − 1 are the closing prices at time t and t − 1, respectively.

3.2. Tests for Randomness

In line with prior literature, we examine randomness in stock returns using a series of alternative test procedures, such as the presence of unit roots, the autocorrelation Q-statistic test, the runs test, and the variance ratio test. In detail, we employ an ADF unit root test to check whether the return series are stationary or not.a By subtracting Yt − 1 from both sides of a first-order autoregressive model we have:

Yt−Yt−1=ρYt−1−Yt−1+ut=(ρ−1)Yt−1+ut

image(13.2)
If ∆Yt = Yt − Yt − 1 and δ = ρ − 1, we then have ∆Yt = δYt − 1 + ut, where δ = 0 and ρ = 1 indicates that the time series follow a random walk. Alternatively, if δ < 0 or ρ < 1, the time series is stationary. Statistical significance is measured using the tau statistic, as suggested by Dickey and Fuller (1981), for testing for the existence of a unit root.b The presence of autocorrelation is then tested by calculating the correlation coefficient of current stock returns and their lags using the mathematical notation:

ρk=∑t=1N−k(rt−r¯)(rt+k−r¯)∑t=1N(rt−r¯)2

image(13.3)
where ρk is the correlation coefficient of the stock return at lag k, N is the number of observations, rt is the stock return in period t, rt + k is the stock return in period t + k, and r¯image is the mean of the stock returns. If ρk is approximately equal to zero, stock returns are normally distributed with a zero mean, indicating independence in the time series for returns. Alternatively, if ρk is different from zero, there is a relation between current stock returns and lagged returns, violating the conditions of the random walk hypothesis. To test the joint hypothesis that all correlation coefficients equal zero, we employ the Ljung–Box portmanteau statistic Q (Ljung and Box, 1978), calculated as

QLB=N(N+2)∑j=1kpj2N−j

image(13.4)
where QLB is the Ljung–Box statistic Q following a chi-square distribution, ρj is the correlation coefficient term at lag jth, and N is the number of observations. If the value of Q exceeds the value of chi-square for k degrees of freedom, there should be at least one ρj value indifferent to zero, rejecting the null hypothesis of randomness in the time series.
The last two alternative test procedures that we employ for the assessment of randomness are the runs test (Bradley, 1968) and the variance ratio (Lo and MacKinlay, 1988; Wright, 2000). The former test is a nonparametric test for examining possible autocorrelation. In general, a run is a sequence of the same movements or signs. By counting the number of runs in the selected sample the existence of autocorrelation can be detected, with a large number of runs indicating a random walk. Under the null hypothesis that stock returns are independent, the number of runs should follow a normal distribution with mean and variance calculated as

E(R)=2n1n2n1+n2+1

image(13.5)

σR2=2n1n2(2n1n2−n1−n2)(n1+n2)2(n1+n2−1)

image(13.6)
where E(R) is the statistical mean, σR2image is the variance of runs, n1 is the number of returns with a positive sign (+), n2 is the number of returns with a negative sign (−), and R is the number of runs. In line with prior literature, we assume n1 > 10 and n2 > 10. Following the estimation of the mean and variance of the runs, our decision rule is to accept the null hypothesis (at a 95% level of significance) if the number of runs (R) stays within the interval of Prob[E(R) − 1.96σR ≤ R  ≤ E(R) + 1.96σR] = 0.95.
The variance ratio test is employed to assess whether or not stock market returns are independently and identically distributed with a constant mean and finite variance.c If the stock returns follow a random walk, the variance of the kth difference at a specific period would be equal to the k time variance of that period’s first difference. We calculate the variance ratio at lag k by multiplying 1/k by the ratio of variance at the kth difference to the variance at the first difference, algebraically formulated as

V(k)=1k×[var (xt+xt−1+…+xt−k+1)]/[var (xt)]            =1k×[var (yt+yt−k)]/[var (yt−yt−1)]            =1+2∑i=1k−1(k−i)k×ρi

image(13.7)
where ρi is the autocorrelation coefficient at lag i for xt. Thus, if stock returns follow a random walk, ρi must equal 0 and the variance ratio for any lag must be equal to 1. We test the null hypothesis of H0: ρi = 0 for every value of i by considering the following equation:

VR(k)=[σ^2(k)]/[σ^2(1)]

image(13.8)
where σ^2(1)image is the variance of the return in period 1 and σ^2(k)image is the variance of the return in period k. We estimate σ^2(1)image as:

σ^2(1)=(T−1)−1∑t=1T(xt+μ^)2=(T−1)−1   ∑t=1T(yt−yt−1−μ^)2

image(13.9)
where μ^image is the sample mean (μ^=T−1∑t=1Txtimage). Similarly, we calculate σ^2(k)image for overlapping long-horizon returns as:

σ^2(k)=m−1∑t=1T(xt+xt−1+…+xt−k+1−k μ^)2=m−1∑t=1T(yt−yt−k−k μ^)2

image(13.10)
where m = (T − k + 1)(1 − kT−1).
To account for possible deviation of stock market returns from the normal distribution assumption, we adopt the rank-based statistical procedure introduced by Wright (2000),d with the rank-based statistic calculated as:

R1(k)=1Tk∑t=kT(r1t+⋯+r1t−k+1)21T∑t=1Tr1t2−1× (k)−1/2

image(13.11)

R2(k)=1Tk∑t=kT(r2t+⋯+r2t−k+1)21T∑t=1Tr2t2−1× (k)−1/2

image(13.12)
where r1t=r(yt)−(T+1)/2(T−1)(T+1)/12image, r2t= −1[r(yt)/(T+1)]image, T is the number of observations of first differences of yt,  timage is the asymptotic variance, r(yt) is the rank of yt, and  −1image is the inverse of the  image-standard normal cumulative distribution function.

3.3. Structural Breaks, Long Memory Dynamics, and Test for Volatility

SEA frontier markets have been subjected to major political and economic policy changes in the past few decades. In order to identify these structural changes (eg, such as the tax reforms, banking sector reforms, crisis and regime shifts), we use the multiple structural break model of Bai and Perron (1998,  2003). They consider the following multiple linear regression with m breaks:

yt=Xβ'+Zt'δj+ɛt  t=Tj−1+1,…,Tj  for j=1,…,m+1

image(13.13)
In this model yt is the dependent variable, xt (p*1) and zt (q*1) are vectors of covariates, β and δj are corresponding vectors of coefficients, and ɛtimage is an error term. The break points (T1…Tm) are treated as unknown, while, conventionally, T0 = 0 and Tm + 1 = T. As β is not subject to structural change whilst δj is, the model is a pure structural break model when p = 0 and a partial structural break model when otherwise. For a specific set of m breakpoints we have {T}m = (T1,…,Tm), where the sum of squared residuals is minimized by

S(β,δ|{T})=∑j=0m{∑t=TjTj+1−1yt−Xt'β−Zt'δj}

image(13.14)
To obtain estimates of (β, δ), we use a least squares regression. In order to select the number of breaks, the Bai–Perron testing procedure first considers the F-statistic to test the null of no structural breaks (m = 0). It then considers the UDmax and WDmax tests, both testing the null of no structural breaks against an unknown number of breaks, given some upper bound M. We adopt the most general specification (eg, trimming point 0.15 and a maximum of three breaks) to allow for all features that the Bai–Perron model offers.
Due to globalization, policy makers in multinational and transnational companies face new challenges in managing their global financial resources so that countries can take full advantage of these opportunities while reducing the potential risk. To tackle this issue, volatility forecast plays a vital role. We applied GARCH (generalized autoregressive conditional heteroscedasticity) family models to investigate the volatility of SEA frontier markets. The volatility modeling process generates mean and conditional variance equations for the series being investigated. Generally, a standard ARIMA (autoregressive integrated moving average) model or a regression model is used to generate the mean equation for the analysis of volatility. Whichever method is used, it includes a term (ɛt) to represent error or residual over time. In this study, we apply an ARIMA model to generate the mean equation for the adopted volatility model. The ARIMA(p, q) process considers linear models of the form shown in Eq. 13.15:

Zt=μ+θ1Zt−1+θ2Zt−2+…+θpZt−p−ϕ1ɛt−1−ϕ2ɛt−2−…−ϕqɛt−q+ɛt

image(13.15)
where ɛt, ɛt − 1,… are present and past forecast errors, and μ, θ1, θ2, …, ϕ1, ϕ2… are parameters to be estimated. The notation Zt is used for the stationary data at time t. When differencing has been used to generate stationarity, the model is said to be integrated and is written as ARIMA(p, d, q), in which p and q represent the order of the autoregressive terms and moving average, respectively. The middle parameter d is simply the number of times that the series needs to be differenced before trend stationarity is achieved.
Engle (1982) presents a basis for formal theory of volatility modeling. At the root of volatility modeling is the distinction between conditional (stochastic) and unconditional (constant) errors. The conditional variance of the error terms is denoted by σt2image and is time varying. Volatility modeling involves adding a variance equation to the original mean equation, which in turn models the conditional variance. Engle (1982) introduces the ARCH (autoregressive conditional heteroscedasticity) model. The ARCH(p) models conditional variance as

σt2=ω+∑i=1pαiɛt−i2

image(13.16)
where ω > 0 and αi > 0.
Many modifications of the basic ARCH(p) model have been developed over time. One of the most widely used volatility models is the GARCH (referenced earlier) developed by Bollerslev (1986). Unlike ARCH, it models the conditional variance as

σt2=ω+∑i=1pαiɛt−i2+∑j=1qβjσt−j2

image(13.17)
where ω > 0 and αi ≥ 0 and βj ≥ 0 to eliminate the possibility of a negative variance.e The GARCH specification in Eq. 13.17 allows for the conditional variance to be dependent on past information. It is explained by past short-run (αi) shocks represented by the lag of the squared residuals (ɛi2)image obtained from the mean equation and by past longer-run (βj) conditional variances (σj2)image. This is referred to as the GARCH(p, q) process. In GARCH models, ∑i=1pαi+∑j=1qβjimage should be less than unity to satisfy stationarity conditions. If βj are all zero, the equation reduces to the ARCH(p) process described in Eq. 13.16—the earliest form of the volatility model developed by Engle (1982). It is rare for the order (p, q) of a GARCH model to be high, while a GARCH(p, q) can be extended to allow for the inclusion of exogenous or predetermined regressors (z) in the variance equation, mathematically notated as

σt2=ω+∑i=1pαiɛt−i2+∑j=1qβjσt−j2+Zt' π

image(13.18)
The EGARCH(p, q) model (exponential GARCH) of Nelson (1991) can also accommodate asymmetric effects, therefore solving important shortcomings of hitherto symmetric models. This is done by specifying the conditional variance in the manner:

loge(σt2)=ω+∑i=1p(αiɛt−iσt−i+γɛt−iσt−i)+∑j=1qβjloge(σt−j2)

image(13.19)
Note that the left-hand side of equation is the logarithm of the conditional variance. This indicates that the leverage effect is exponential; hence, guaranteeing that the forecasts of the conditional variance will be nonnegative. One reason that EGARCH has been popular in financial applications is that the conditional variance, σt2image, is an exponential function, thereby removing the need for a constraint in the parameters to ensure a positive conditional variance (Longmore and Robinson, 2005). The model also permits asymmetries via the γ term. The presence of leverage effects is tested by the hypothesis that γ < 0. If γ < 0, negative shocks increase volatility, while if γ = 0, the model is symmetric. Two additional advantages from using the EGARCH family models are that the values of p and q are very rarely high, and that these models tend to be parsimonious. Furthermore, before generating an optimal model for our return series, we also test possible misspecification using the Ljung–Box Q-statistic. The Q-squared (QSQ) statistic is employed to check the ARCH in the residuals, while if more than one volatility model with significant parameters is found, the model with the maximum log likelihood (LL) criterion is selected as the most optimum.

4. Empirical Results

4.1. Results From Randomness Tests

Panels A and B in Table 13.2 present the results of the augmented Dickey–Fuller (ADF) unit root examination using three alternative specifications: (1) with intercept, (2) with intercept as a linear trend, and (3) without intercept.

Table 13.2

Unit Root Test Results for the Index of Each SEA Frontier Stock Market

Panel A: daily returns
Index VN Index HNX Index LSX Index CSX Index
Intercept −21.103*** −38.943*** −16.344*** −23.526***
Intercept, linear trend −21.126*** −38.949*** −16.337*** −23.568***
None −21.076*** −38.952*** −16.346*** −23.311***
Panel B: weekly returns
Index VN Index HNX Index LSX Index CSX Index
Intercept −15.347*** −17.273*** −10.354*** −11.071***
Intercept, linear trend −15.367*** −17.281*** −10.363*** −11.010***
None −15.316*** −17.292*** −10.402*** −10.930***

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed on the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively. *** signifies the rejection of the null hypothesis of having a unit root at a 1% level of significance.

According to the results, we have to reject the existence of a unit root for all specifications and for all daily (Panel A) and weekly (Panel B) return time series. The values of ρ are all consistently negative and significant at the 1% level, irrespective of the model applied. On this basis, we have to reject the null hypothesis of stock market returns following a random walk for the countries of Vietnam, Laos, and Cambodia. This finding corroborates prior literature on the informational efficiency of the Vietnamese market and it is very similar to the ADF unit root test results of Guidi and Gupta (2011). A very similar picture is portrayed by the autocorrelation tests presented in Table 13.3. As mentioned in the previous section, the autocorrelation Q-statistic test of Ljung and Box (1978) examines the null hypothesis of no autocorrelation among return series. According to both Panels A and B, there is a statistically significant autocorrelation for all return series and up to 15 lags for the stock markets of Vietnam and Laos, but surprisingly not for the case of Cambodia. The strongest autocorrelation (AC) is observed in the case of the VN Index for both the daily and weekly returns (for lag = 1, AC values of 0.30 and 0.16 with a Q-statistic of 308.51 and 20.07, respectively). All values are significant at the 1% level. In contrast, apart from the case of the weekly return series for the Cambodian market using one lag, which is positive and significant at the 5% level, all other values are statistically insignificant for all selected lags. This finding suggests that the Cambodian stock market is weak-form efficient. However, the presence of randomness and informational efficiency in this case should always be assumed with caution, as the market is relatively new with a limited number of securities listed on it.

Table 13.3

Autocorrelation Test Results for Daily SEA Frontier Stock Market Indices

Panel A: daily returns
Index VN Index HNX Index LSX Index CSX Index
Lag AC Q-stat AC Q-stat AC Q-stat AC Q-stat
1 0.30 308.51*** 0.18 72.16*** 0.18 30.80*** 0.056 1.92
3 0.02 318.70*** 0.04 77.36*** 0.04 39.85*** 0.017 3.52
5 0.12 395.94*** 0.07 101.01*** 0.09 52.20*** 0.071 6.84
7 0.05 430.39*** −0.01 101.30*** 0.02 58.16*** 0.003 7.35
9 0.03 436.23*** 0.05 109.57*** 0.04 68.00*** −0.009 7.52
11 0.05 451.55*** 0.02 110.60*** 0.03 70.09*** 0.082 12.62
13 0.04 458.43*** 0.04 115.25*** −0.03 70.99*** 0.024 18.42
15 0.06 487.26*** 0.01 120.39*** −0.08 80.12*** 0.055 20.35
Panel B: weekly returns
Index VN Index HNX Index LSX Index CSX Index
Lag AC Q-stat AC Q-stat AC Q-stat AC Q-stat
1 0.16 20.07*** 0.21 20.61*** 0.35 21.32*** 0.16 3.86**
3 0.11 47.56*** 0.09 28.63*** −0.08 22.57*** −0.03 4.56
5 0.16 71.56*** 0.06 36.77*** −0.09 24.53*** 0.11 6.82
7 0.07 76.29*** 0.06 39.47*** 0.01 25.19*** −0.04 7.05
9 0.06 78.91*** 0.03 39.82*** 0.03 25.53*** −0.08 8.18
11 0.00 79.14*** −0.01 39.89*** 0.07 26.90*** 0.02 8.53
13 −0.02 80.12*** −0.04 41.81*** 0.04 28.47*** −0.06 11.04
15 −0.11 88.63*** −0.07 45.34*** −0.04 28.92** 0.02 11.29

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed on the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively. *** and ** signify the rejection of the null hypothesis of no autocorrelation at the 1 and 5% levels of significance, respectively.

Overall, these results are consistent to those presented in Aumeboonsuke (2012), Vinh and Thao (2013), and Phan and Zhou (2014) for the Vietnamese stock market, and the first-ever empirical evidence reported for the stock markets of Cambodia and Laos.
Testing possible randomness using the nonparametric runs test and the variance ratio further strengthens these findings. According to Panel A in Table 13.4, the runs test report z-statistics of −11.984 and −4.677 for the two indices of VN and HNX are statistically significant at the 1% level for the daily data. The negative values in this case indicate that daily upward (downward) return movements are followed by subsequent upward (downward) movements, violating the assumption of randomness. Hence, the results suggest that the Vietnamese stock market exhibits clear evidence of autocorrelation, rejecting the null hypothesis of randomness, and further corroborating the prior findings of Guidi and Gupta (2011), Vinh and Thao (2013), and Phan and Zhou (2014). On the contrary, the results for Laos suggest no presence of autocorrelation in the return series (z-statistic of −0.480), while for the Cambodian stock market the null hypothesis of a random walk is rejected at the 5% level (z-statistic of −2.178).

Table 13.4

Runs Test Results for Daily SEA Frontier Stock Market Indices

Panel A: daily returns
Index VN Index HNX Index LSX Index CSX Index
E(R) 1715.765 1097.482 473.230 259.699
var(R) 856.7655 547.981 226.839 108.577
StDev(R) 29.271 23.409 15.061 10.420
Z-stat −11.984*** −4.677*** −0.480 −2.178**
Panel B: weekly returns
Index VN Index HNX Index LSX Index CSX Index
E(R) 370.903 231.499 82.657 68.586
var(R) 184.652 114.998 39.203 32.376
StDev(R) 13.589 10.724 6.261 5.689
Z-stat −4.703*** −3.497*** −0.424 −0.454

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed in the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively. *** signifies the rejection of the null hypothesis of no autocorrelation at a 1% level of significance.

In the case of the weekly data (Panel B), the runs test further confirms the previous findings with a z-statistic of −4.703 and −3.497 for the Vietnamese market. Nonetheless, the test fails to reject the null hypothesis of a random walk in the cases of Laos and the Cambodian stock market indices.
To shed more light into these surprising findings, we assess their robustness using the more rigorous variance ratio test of Wright (2000). All results for daily and weekly returns for the four indices are presented in Table 13.5 in Panels A and B, respectively.

Table 13.5

Variance Ratio Test Results for Daily SEA Frontier Stock Market Indices

Panel A: daily returns
Index VN Index HNX Index LSX Index CSX Index
Ranks 19.032*** 17.737*** 14.533*** 11.606***
Rank scores 20.559*** 18.765*** 14.701*** 12.246***
Signs 12.943*** 12.944*** 10.313*** 8.297
Panel B: weekly returns
Index VN Index HNX Index LSX Index CSX Index
Ranks 11.636*** 8.484*** 5.574*** 4.953***
Rank scores 12.708*** 8.889*** 5.421*** 5.467***
Signs 7.982*** 5.875*** 4.320*** 2.799**

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed in the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively. *** and ** signify the rejection of the null hypothesis of random walk at the 1 and 5% levels of significance, respectively.

According to Panel A, all rank-based test results and rank scores are positive and statistically significant at the 1% level, rejecting the null hypothesis of weak-form efficiency in the examined SEA frontier markets. As regards the sign-based results, the tests on the daily returns further confirm what had been reported previously in the case of Vietnam and Laos, but provide conflicting results from the Cambodian stock market, as the S-value reported is 8.297, which is insignificant in statistical terms.
The results from the examined weekly returns presented in Panel B further confirm the case of informational inefficiency for the examined markets, in which the null hypothesis of random walks is rejected in both the rank-based and sign-based tests. All values are positive and significant at the 1% level. This conclusion supports prior findings in the literature (Guidi and Gupta, 2011; Vinh and Thao, 2013; Phan and Zhou, 2014).

4.2. Results on Structural Breaks and Volatility

The results from Bai and Perron (2003) structural break test for the daily and weekly return series are presented in Table 13.6. Three break dates are identified in each series, and these are statistically significant.f These break dates are then incorporated as variance regressors in the modeling of stock market volatility.

Table 13.6

Bai-Perron (2003) Structural Break Test Results

Panel A: daily returns
Country Global L breaks versus none
F-stat and scaled F-stat Weighted F-stat UDMax stat** WDMax stat** Break dates
VN Index

6.635

4.983

5.275

6.635

5.922

7.594

 6.635 7.594

Oct. 27, 2003

Feb. 28, 2007

Jan. 4, 2009
HNX Index

5.592

4.816

4.063

5.592

5.723

5.849

 5.592 5.849

Oct. 17, 2007

Feb. 25, 2009

Jun. 17, 2010
LSX Index

0.352

3.467

2.981

0.352

4.120

4.291

 3.467 4.291

Feb. 20, 2012

Jan. 22, 2013

Jan. 27, 2014
CSX Index

1.857

3.427

2.031

1.857

4.072

2.924

 4.073 3.427

Mar. 16, 2012

Mar. 22, 2013

Feb. 08, 2014
Panel B: weekly returns
Country Global L breaks versus none
F-stat and scaled F-stat Weighted F-stat UDMax stat** WDMax stat** Break dates
VN Index

5.972

4.185

4.17

5.972

4.973

5.927

 5.972 5.972

Oct. 17, 2003

Jan. 19, 2007

Mar. 6, 2009
HNX Index

4.434

3.956

3.215

4.434

4.698

4.629

 4.434 4.698

Oct. 19, 2007

Feb. 13, 2009

Jun. 18, 2010
LSX Index

0.684

3.147

4.164

0.685

3.740

5.995

 4.164 5.995

Jul. 22, 2011

Jan. 27, 2012

Feb. 1, 2013
CSX Index

2.922

3.049

2.002

2.922

3.623

2.882

 3.049 3.624

Sep. 21, 2012

May 31, 2013

Oct. 25, 2013

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed in the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively.

The mean and conditional variance equations for daily and weekly series are presented in Tables 13.7 and 13.8, respectively. The choice of ARIMA for the data is based on the model (1) being parsimonious, (2) having significant parameters, (3) having errors that are white noise, and (4) reporting a minimum Schwarz Bayesian criterion (SBC) (Schwarz, 1978). The mean equations act as a basis for generating the conditional variance equations for each series. To obtain the optimal GARCH(p, q) model, all combinations of (p) = (0,1,2) and (q) = (0,1,2) were considered (except for p = q = 0), as suggested by Angelidis et al. (2004). The threshold order determines the impact (or otherwise) of news shocks. The threshold order of zero means that the volatility model is symmetric; for example, the impact of good news equals the impact of bad news in terms of volatility effect. A threshold order of one means the model is asymmetric, that is, the impact of good news does not equal the impact of bad news. All combinations of symmetric and asymmetric volatility models were run. In most instances more than one of the ARCH, GARCH, EGARCH, and/or PGARCH models with significant parameters are found. The model with maximum LL criterion is selected as the optimal model for each series, and according to the QSQ statistic all models presented are correctly specified (QSQ > 0.05). According to our results, the dummy variable(s) which address(es) structural break(s) is (are) insignificant in most of the cases. For example, according to Table 13.7 Panel B, a significant D2 is reported only for the case of the Vietnam’s VN Index (z-value of 2.320, significant at the 5% level). Thus, these findings generally indicate that although there are structural breaks in the return series, they do not significantly influence the volatility of the examined markets.

Table 13.7

Results From Volatility Models (Daily Returns)

Panel A: mean equation
VN Index AR(1) MA(1) MA(2)

ARIMA(1,0,2)(0,0,0)

z-statistic

0.994

452.870***

−0.745

−39.238***

−0.237

−12.513***
HNX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

0.669

5.989***

−0.556

−4.648***
LSX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

0.869

22.023***

−0.887

−26.312***
CSX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

0.957

52.092***

−0.982

−102.824***
Panel B: conditional variance equations
ω α1 β1 D2
VN Index

EGARCH (1,1)

z-statistic

−0.789

−10.493***

LL = 10,329.73

0.434

14.101***

SBC = −6.007

0.949

131.426***

QSQ(12) = 32.192

(0.650)

0.263

2.320**
HNX Index

EGARCH (1,1)

z-statistic

−0.651

−7.235***

LL = 5,688.361

0.397

10.995***

SBC = −5.173

0.956

102.933***

QSQ(12) = 4.616 (0.948)
LSX Index

EGARCH (1,1)

z-statistic

0.000

3.861***

LL = 3,036.301

0.361

5.524***

SBC = −6.155

0.373

3.531***

QSQ(12) = 26.065 (0.889)
CSX Index

EGARCH (1,1)

z-statistic

−1.293

−2.970***

LL = 1,828.873

0.370

4.519***

SBC = −5.941

0.879

19.312***

QSQ(12) = 40.825 (0.267)

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed in the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively. *D2 is a dummy variable representing structural breaks. *** and ** signify the rejection of the null hypothesis of random walk at the 1 and 5% levels of significance, respectively.

Table 13.8

Results From Volatility Models (Weekly Returns)

Panel A: mean equation
VN Index AR(1) AR(2) MA(1) MA(2)

ARIMA(2,0,2)(0,0,0)

z-statistic

−0.360

−2.303**

−2.303

4.028***

0.520

2.944***

−0.455

−2.624***
HNX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

0.669

5.990***

−0.556

−4.648***
LSX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

−0.943

−63.156***

0.983

223.576***
CSX Index AR(1) MA(1)

ARIMA(1,0,1)(0,0,0)

z-statistic

0.544

2.540**

−0.530

−2.678**
Panel B: conditional variance equations
ω α1 β1 D1 D2 D3
VN Index

EGARCH (1,1)

−0.585

−4.101***

0.407

5.625***

0.959

62.687***

0.685

3.075***

−0.455

−2.624***
z-statistic LL = 1426.6 SBC = −3.8

QSQ(12) =  27.2

(0.854)
HNX Index
EGARCH (1,0)

−6.259

−46.159***

0.360

4.193***

−5.021

−15.582***

1.283

3.968***
z-statistic LL = 725.3 SBC = −3.1 QSQ(12) =  19.8 (0.975)
LSX Index
EGARCH (1,1)

−0.557

−5.24***

−0.409

−5.460***

0.891

68.347***
z-statistic LL = 407.3 SBC = −4.7 QSQ(12) =  45.8 (0.128)
CSX Index
EGARCH (1,0)

−7.530

−28.748***

0.517

2.541**

−1.383

−2.362**

−4.494

−9.045***

1.641

2.438**
z-statistic LL = 310.1 SBC = −4.2

QSQ(12) =  10.3

(0.999)

Notes: The VN Index represents large-cap stocks (>VND 120 billion), while the HNX Index consists of medium and small-cap companies listed in the Vietnamese stock exchange. The LSX Index and the CSX Index are the composite indices of the Laos Securities Exchange and the Cambodia Stock Exchange, respectively.

* D1, D2, and D3 are dummy variables represent the structural breaks. *** and ** signify the rejection of the null hypothesis of random walk at the 1 and 5% levels of significance, respectively.

The empirical results reveal that EGARCH volatility models are optimal for all cases except the daily LSX Index. This supports the findings of Alberg et al. (2008), who investigated the forecasting performance of various volatility models and concluded that the EGARCH volatility model generates better results. The analyses also show that βi > αi in all daily series. This indicates that there is a relatively long-term impact of shocks on stock markets. Natural disasters, political unrest, unstable economic situations, and decisions concerning macroeconomic fundamentals create these longer effects on stock markets. These results are also consistent with the weekly data. However, the short-term impact of shocks (αi > βi) is found in the cases of the weekly HNX Index and the CSX Index. Symmetric volatility models are statistically significant in both daily and weekly series, implying that the impact of positive and negative news or shocks is the same in magnitude. This indicates that local news announcements (good and bad) regarding macroeconomic fundamentals, politics, and specifics about companies have the same impact on daily and weekly returns. Consequently, on the basis of these findings, we have to reject the null hypothesis of randomness in the return time series.

5. Conclusions

In this study, a number of econometric procedures are employed to investigate the weak-form efficiency of the three SEA frontier markets of Vietnam, Laos, and Cambodia. According to our results, the earliest and largest market of the three, that is Vietnam—is found to be weak-form inefficient, regardless of the test procedure used. On the other hand, in the case of the LSX there are mixed results. For example, our null hypothesis of a random walk is rejected using unit root, autocorrelation, and variance ratio tests, but it is accepted using a runs test on both daily and weekly data. In the case of the CSX the results are particularly interesting, as the market appears to be weak-form efficient according to the results from the autocorrelation Q-statistic test (using both daily and weekly data), runs tests (weekly data), and the variance ratio test (daily data). Hence, this study brings about a surprising finding: that this younger and considerably smaller SEA frontier market shows some positive signs of weak-form efficiency. Nonetheless, none of the three markets examined demonstrates consistently significant evidence on weak-form efficiency across all methodologies employed.
As regards the empirical results on market volatility and the long memory dynamics of these markets, the EGARCH volatility models appear to be optimal for all cases except that of the daily LSX Index. Results also indicate that there is a relatively long-term impact of shocks on these stock markets. For example, natural disasters, political unrest, unstable economic situations, and decisions concerning macroeconomic fundamentals create these longer effects on stock prices. Our results indicate a short-term impact of shocks in the case of the weekly indices for the HNX and the CSX. Moreover, symmetric volatility models are statistically significant in both daily and weekly series, implying that the impacts of positive and negative news or shocks are the same in magnitude.
Any attempt by relevant governments to increase the informational efficiency of their equity markets so as to promote their business environment and attract foreign invested capital requires above all the creation of an appropriate legislative and regulatory framework. For example, some of the key changes urgently needed in this very early stage in the life of these markets are (1) the strengthening of the role and capacities of the relevant securities commissions to raise awareness and confidence in the market across local and foreign investors; (2) completion of the securities laws and the appropriate regulatory framework to improve information transparency and restrict insider trading, and (3) propagation of an open mechanism and incentive schemes to attract more listings of local companies and additional investors (especially foreign and institutional investors) for the purposes of market development, product diversification, and competition enhancement.

References

Alberg D, Shalit H, Yosef R. Estimating stock market volatility using asymmetric GARCH models. Appl. Financ. Econ. 2008;18(15):1201–1208.

Angelidis T, Benos A, Degiannakis S. The use of GARCH models in VaR estimation. Stat. Methodol. 2004;1(2):105–128.

Aumeboonsuke V. Weak-form efficiency of six equity exchanges in ASEAN. Eur. J. Sci. Res. 2012;84(4):532–538.

Bai J, Perron P. Estimating and testing linear models with multiple structural changes. Econometrica. 1998;66(1):47–78.

Bai J, Perron P. Computation and analysis of multiple structural change models. J. Appl. Economet. 2003;18(1):1–22.

Bollerslev T. Generalized autoregressive conditional heteroscedasticity. J. Econom. 1986;31:307–327.

Borges MR. Random walk tests for the Lisbon stock market. Appl. Econ. 2011;43(5):631–639.

Bradley JV. Distribution-Free statistical tests. Manage. Sci. 1968;16(1):141–143.

Cowles A. A revision of previous conclusions regarding stock price behavior. Econometrica. 1960;28(4):909–915.

Dickey D, Fuller WA. Likelihood ratio statistics for autoregressive time series with a unit root. Econometrica. 1981;49(4):1057–1072.

Engle RF. Autoregressive conditional heteroskedacticity with estimates of the variance of United Kingdom inflation. Econometrica. 1982;50(4):987–1007.

Fama EF. The behavior of stock market price. J. Bus. 1965;38:34–105.

Fama EF. Efficient capital markets: a review of theory and empirical work. J. Financ. 1970;25(2):283–417.

Fama EF, French KR. Permanent and temporary components of stock prices. J. Polit. Econ. 1987;96(2):246–273.

Fama EF, French KR. Dividend yields and expected stock returns. J. Financ. Econ. 1988;22(1):3–25.

Guidi, F., Gupta, R., 2011. Are ASEAN Stock Market Efficient? Evidence From Univariate and Multivariate Variance Ratio Tests. Discussion Papers in Finance, Griffith University.

Kendall MG. The analysis of economic time-series. Part I: prices. J. Roy. Stat. Soc. 1953;116(1):11–34.

Ljung GM, Box GEP. On a measure of lack of fit in time series models. Biometrika. 1978;65(2):297–303.

Lo AW, MacKinlay AC. Stock market prices do not follow random walks: evidence from a simple specification test. Rev. Financ. Stud. 1988;1(1):41–66.

Loc, T.D., 2006. Equitisation and Stock Market Development: The Case of Vietnam. Unpublished PhD Thesis, University of Groningen.

Longmore R, Robinson W. Modeling and forecasting exchange rate dynamics in Jamaica: an application of asymmetric volatility models. Money Aff. 2005;18(1):23–56.

Mishra P. Weak-form market efficiency: evidence from emerging and developed world. J. Commer. 2012;3(2):26–34.

Nelson DB. Conditional heteroscedasticity in asset returns: a new approach. Econometrica. 1991;59(2):347–370.

Nelson DB, Cao CO. Inequality constraints in the univariate GARCH model. J. Bus. Econ. Stat. 1992;10(2):229–235.

Nisar S, Hanif M. Testing weak-form of efficient market hypothesis: empirical evidence from South-Asia. World Appl. Sci. J. 2012;17(4):414–427.

Osborne MFM. Periodic structure in the Brownian motion of stock prices. Oper. Res. 1962;10(3):345–379.

Pearson K. The problem of the random walk. Nature. 1905;72:294.

Phan KC, Zhou J. Market efficiency in emerging stock markets: a case study of the Vietnamese stock market. IOSR J. Bus. Manage. 2014;16(4):61–73.

Righi M, Ceretta P. Random walk and variance ratio tests for efficiency in the sub-prime crisis: evidence for the U.S. and Latin markets. Int. Res. J. Finance Econ. 2011;72:25–32.

Samuelson PA. Proof that properly anticipated prices fluctuate randomly. Ind. Manage. Rev. 1965;6(2):41–49.

Schwarz G. Estimating the dimension of a model. Ann. Stat. 1978;6(2):461–464.

Sharma JL, Kennedy RE. A comparative analysis of stock price behavior on the Bombay, London, and New York stock exchanges. J. Financ. Quant. Anal. 1977;12(3):391–413.

Smith G. The changing and relative efficiency of European emerging stock markets. Eur. J. Finance. 2012;18(8):689–708.

Suleman MT, et al. Testing the weak-form of efficient market hypothesis: empirical evidence from Asia-Pacific markets. Int. Res. J. Finance Econ. 2010;58:121–133.

Szakmary A, et al. Filter tests in Nasdaq stocks. Financ. Rev. 1999;34(1):45–70.

Tsai H, Chan KS. A note on inequality constraints in the GARCH model. Economet. Theor. 2008;24(3):823–828.

Vinh V.X., Thao, L.D.B., 2013. Empirical investigation of efficient market hypothesis in Vietnam Stock market. SSRN Electronic Journal 03/2013.

Worthington, A., Higgs, H., 2003. Tests of Random Walks and Market Efficiency in Latin American Stock Markets: an Empirical Note. School of Economics and Finance Discussion Papers and Working Papers Series, No. 157. Queensland University of Technology, School of Economics and Finance, Brisbane, Australia.

Worthington A, Higgs H. Random walks and market efficiency in European equity markets. Global J. Finance Econ. 2004;1(1):59–78.

Worthington, A., Higgs, H., 2005. Weak-Form Market Efficiency in Asian Emerging and Developed Equity Markets: Comparative Tests of Random Walk Behavior. School of Accounting and Finance Working Paper Series, No. 05/03. University of Wollongong, Wollongong, Australia. Available from: http://ro.uow.edu.au/commpapers/199

Wright JH. Alternative variance-ratio tests using ranks and signs. J. Bus. Econ. Stat. 2000;18(1):1–9.

a A time series process is characterized as stationary if its mean and variance are constant and the covariance between two different time periods depends upon the lag between them. The mean and variance of the time series are estimated as (Yt) = μ and Var(Yt) = E(Yt–μ)2, respectively. The covariance of Y values at times t and t + k is estimated as γk = E[(Yt–μ)(Yt+k–μ)], while a unit root is defined by using the first-order autoregressive model of Yt = ρYt − 1 + μt with (−1 ≤ ρ ≤ 1). If ρ = 1, there is a unit root in the time series process.

b Dickey and Fuller (1981) show that the coefficient after transformation to δ does not follow a normal distribution when the size of the sample is large. This results in conventional t-statistic results being erroneous.

c This will indicate that the variance is a linear function of the relevant holding periods.

d Wright (2000) has extended the Lo and MacKinlay test (1988) by introducing rank differences, so that the variance ratio test could be applied in cases in which the distribution of returns is not normally distributed.

e Nonetheless, prior literature suggests that in practice this βj ≥ 0 constraint can be overrestrictive (Nelson and Cao, 1992; Tsai and Chan, 2008).

f These breaks in the return time series are triggered by important political or economic events in the countries in question. For example, the structural break of Oct. 27, 2003, for Vietnam follows immediately after the signing of an agreement between the governments of Vietnam and the United States to start—for the first time since the end of the Vietnam War—commercial flights between the two countries. Meanwhile, the break of Feb. 28, 2007, is driven by the Vietnamese government’s announcement to invest US$33bn in infrastructure projects, for example, a high-speed rail link between Hanoi and Ho Chi Minh City.

..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset
3.144.19.243