Financial liberalization, stock market volatility and outliers in emerging economies

Abstract

In this article, we test whether the structure of emerging market volatility has changed and assess the link between the structural changes in volatility behaviour and financial liberalization events. The opening of financial markets tends to generate outlying returns around the opening dates, thus giving the appearance of increases in market volatility. We include outlier detection methodologies in our location of endogenous breaks in order to filter out this effect. Our results suggest that changes in volatility behaviour have indeed been induced by financial liberalization of emerging markets, but the change is not always in the same direction: Latin American countries have enjoyed lower volatility whereas Asian countries seem to have suffered increases in market instability. Additionally, all markets become more subject to occasional large shocks.

Acknowledgements

We thank Elena Andreou and Geert Bekaert for useful comments. We thank participants at seminars at the Universidad de Alicante, Universidad Autónoma de Madrid, Universitat Pompeu Fabra, the International Conference on Computing in Economics and Finance (Amsterdam, 2004), the 10th Meeting of LACEA (Paris, 2005), the Emerging Markets Finance Conference at the Cass Business School (London, 2005) and the 1st Meeting of the Methods in International Finance Network (Maastricht, 2007). Financial assistance from the Plan Especial de Investigacion de la Universidad de Navarra and the Spanish Ministry of Science and Technology through grant SEJ2005-06302/ECON are gratefully acknowledged.

Notes

¹ In this article, we use data from S&P/IFC Composite Index. The database provides a long time-series of information–since 1976:01–for eight countries.

² This last article is especially relevant to our discussion: the authors find many structural changes in volatility, most of which seem to be induced by momentary instability and not by permanent changes in volatility structure.

³ This assumes that, if σ _t ² is a random variable, as in stochastic volatility models, then it is independent from ε _t . If σ _t ² is a deterministic function of information known at time t − 1, as in Generalized Autoregressive Conditional Hetroscedasticity (GARCH) models, then the equality follows immediately.

⁴ Tests are being designed that are more robust to the mean structure of the return process: see, for example, Chen et al. (2005) or Rodrigues and Rubia (2005).

⁵ For example, Aggarwal et al. (1999) found evidence of structural breaks in the variance of all the series they analysed, and sometimes the breaks were very frequent: in the case of Argentina, they find evidence of 10 breaks in a total of 10 years of data (p. 45). This finding is easy to interpret in the light of our discussion, and by looking at the graphs in the original article that represent the returns along with the estimated variances. The breaks are detected by using cumulative sums of squares of returns, so large outlying returns cause the appearance of the break. The authors find that when a dummy variable is included for the whole period until the next break (i.e. until the next big return signals a break) GARCH-type effects disappear. This should be expected given that the effect of the outlying return would be accounted for in the variance equation by the period-by-period dummies.

⁶ We assume that parameters of the mean process are constant across the two regimes. Given that the mean structure of returns is weak–usually φ₁ is not found to be significantly different from 0–this assumption is not too restrictive and it simplifies the interpretation of the break parameter. An analysis that allows for change of the mean parameters was also carried out, with very similar results–estimated values of φ₁ were not significantly different from 0 and the dating of the breaks was consistent.

⁷Tsay (2002) assumes a constant variance of regular innovations , an assumption that is difficult to justify in a financial time series.

⁸ This is the ‘data augmentation’ step of the procedure (Tsay, 2002).

⁹ Assuming a conditional prior for the mean parameters , the posterior, given the data R, is f(φ|R;α,π,δ,D,ε) = [(f(R|φ,α,π,δ,D,ε)·f(φ|α,π,δ,D,ε))/(f(R|α,π,δ,D,ε))].

¹⁰ The number of grid points used in our analysis in Section IV is 2000 for and, given their more constrained range, 200 for and . For π the grid is taken to be the integers in [0.15T, 0.85T]. Between grid values the cumulative probability is assumed to grow linearly.

¹¹ Weekly or daily data are avalaible since 1988 and 1995, respectively.

¹² These indexes, formerly calculated by the International Finance Corporation (IFC), are dollar-denominated price indexes of the stock markets in each country. We use the global index and not the investable, which is a narrower index that is only available from the 1990s onwards. The S&P/IFC global index represents the performance of the most active stocks in each market analysed and attempts to be the broadest possible indicator of market movements, corresponding to at least 75% of total capitalization. For further information on these widely used indexes, consult www.standardandpoors.com

¹³ Data availability and comparability also dictated the final set of countries analysed. Some local indexes, such as Brazil's Bovespa and Chile's IGPA (the general index of weighted share prices), were available for longer periods, but we opted for using a uniformly calculated index to make comparison across countries more meaningful and not subject to the different methodologies used by the countries. Still, one would ideally use as long a series as possible.

¹⁴ We calculate regular returns as . Filtered returns correspond to the outlier-free series .

¹⁵ We use , where is the variance of the residuals of a first-stage AR(1) regression of r_t . Alternative values for ξ² were tried, but the results were not affected: neither the probability nor the estimated size of the outliers changed, as long as ξ² was large enough compared to the unconditional variance of the outlier-free process.