Transmission channels of international financial crises to African stock markets: the case of the euro sovereign debt crisis

ABSTRACT

This article investigates the effects of the European sovereign debt crisis on African stock markets within a Bayesian shrinkage VAR framework. This method allows us to consider both North African and Sub-Saharan African stock markets, and provides a flexible parsimonious specification. The results reveal varying reactions of the impulse response functions. The most exposed African stock markets are those of Egypt, South Africa and Mauritius, while the least affected stock market is, surprisingly, that of Ivory Coast. Our analysis shows that, in addition to direct transmission, several macroeconomic and market channels, such as commodities, exports, and exchange rates, are relevant. Specifically, countries with strong commercial links to European countries will be most impacted by the crisis. The severity of transmission also depends on the country’s dependence on commodities.

KEYWORDS:

• African stock markets
• European Sovereign debt crisis
• transmission effects
• Bayesian vector autoregression

JEL CLASSIFICATION:

• F36
• F49
• E32
• O55

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ See Bernanke, Boivin, and Eliasz (2005).

² VAR models were introduced by Sims (1980). They capture the interdependencies among multiple time series. VAR models have a long and successful tradition in the forecasting literature (Doan, Litterman, and Sims 1984; Litterman 1986) and in monetary policy analysis (Kim 2001; Ehrmann and Fratzscher 2006). However, standard VAR models are rarely applied to more than eight time series (Bernanke, Boivin, and Eliasz 2005; and Feldkircher 2015).

³ See (Bańbura, Giannone, and Lenza 2015).

⁴ The prior on the intercept, c, is diffused.

⁵ For a complete review of the various methods for the settlement of priors, the reader can refer to Koop and Korobilis (2010) and Korobilis (2016). In addition, Koop and Korobilis (2009) present the results of a BVAR model estimated using different priors and argue that different priors provide similar empirical results. Lenza, Primiceri, and Giannone (2010) test different priors in BVARs with different sizes (SMALL, MEDIUM, and LARGE) and recommend the Minnesota priors.

⁶ Litterman (1986) and other researchers at the University of Minnesota developed priors for VAR coefficients, which are known as the Minnesota priors.

⁷ The reader can refer to Litterman (1986) for more details concerning the settlement and hypothesis for each parameter.

⁸ See Kadiyala and Karlsson (1997) for more details.

⁹ As stated by Primiceri (2005) and Copy (2011), the MCMC method delivers smoothed estimates of the parameters of interest based on the entire available set of data. This method also allows the estimation of the function of the parameters, such as an impulse response function, with the uncertainty of the unknown parameters.

¹⁰ Bannigidadmath and Narayan (2016) explain why daily data maybe better than monthly data when the objective is to obtain as much information as possible from data. However, even with monthly, data we found qualitatively the same results.

¹¹ All series are integrated of order one based on standard unit root tests, available on request. This is a standard result in the literature on such financial series.

¹² To select the order of lags, we estimate a VAR model for the number of lags between one and eleven and rely on the AIC and Schwarz information criterion, among others, to determine the optimal number of lags. According to the AIC, FPE, HQ and the SC criterion, the optimal lag number is p = 1. The residuals are found to be white noise when the lag length is 1. This model is estimated by ordinary least squares (OLS). The results are available on request.

¹³ As the dimension increases, we set the tightness of the prior so that the BVAR has the same in-sample fit as the VAR. Thus, we ensure that the in-sample fit is constant and that the Bayesian shrinkage hyperparameter corresponds with the model dimensions. This allows us to compare models of different sizes.

¹⁴ The estimation was implemented using Gibbs sampling procedure. More precisely, we use the MCMC method with 10,000 iterations, the first 2,000 of which were discarded to minimize the effects of initial conditions. To assure convergence of the algorithm, we imposed proper priors on the parameters, as explained in the methodology section. No problems were experienced in achieving convergence, and both alternative starting values and the use of 20,000 iterations produced essentially the same results.

¹⁵ The code to estimate the specifications (VAR and BVAR) and produce the GIRFs was programmed using MATLAB software.

¹⁶ Before presenting our main results, we ensure that the linkages between the variables established by the model are plausible and statistically significant, which may not be the case with model overfitting.

¹⁷ We opt for the GIRFs because they are invariant to the ordering of variables and do not require identification restrictions.

¹⁸ This method is used by the most widely used data sources in economics, the National Income and Product Accounts (NIPAs) from the U.S. Bureau of Economic Analysis. This method allows us to draw the missing points by computing a piecewise polynomial function connecting two points.