Foreign exchange market bid-ask spreads and the market price of social unrest.
The BID-ASK spread in the foreign exchange market represents equilibrium prices at which banks buy and sell currencies. Foreign exchange market researchers have focused on the behavior of spot and forward exchange rates with very little attention devoted to the issue of equilibrium pricing of the spread. The little research that has been published largely describes the statistical properties of spreads for industrial countries (e.g., Boothe, 1988; and Glassman, 1987). The fact that industrial country spreads change little through time probably contributes to a lack of interest in the dynamics of spreads. For instance, Goodhart and Figliuoli (1991) study the spreads on industrial country currencies against the US dollar and conclude:
We have, therefore, no evidence here that posted spreads in the forex market are sensitively responsive to changes in market conditions, although it does appear that in some cases the conventional spread for a currency may vary from market to market (e.g., from Europe to North America), and thus be time dependent. Instead, it is perhaps more likely that banks will respond to changing perceived risk by shifting the amounts in which they will subsequently be prepared to deal over the telephone. [p. 29]
The fact that Goodhart and Figliuoli find no evidence of a time-varying spread over three days for industrial currencies is, perhaps, to be expected. Their analysis was focused on the intra-daily movements in spot rates, so the spread was simply a minor aside in their paper. Yet we should not consider the spread to be a constant in cases where the market receives important information that changes the expected future profitability associated with taking positions in foreign exchange. Such events are often difficult to identify in large, well-diversified industrial countries, but for developing countries that are subject to recurring social or political turmoil, we may expect to witness large swings in the bid-ask spread on domestic currency.
We address the issue of the foreign-exchange market price of social unrest by relating changes in the foreign exchange market bid-ask spread, an important cost of transacting, to such unrest. Section 4 contains empirical tests of the model over time for the rand in South Africa - a country that has been subjected to easily identifiable events that should have had an impact on expected profitability from foreign exchange positions. Our model predicts that changing country risk in South Africa can be reflected in changing spreads on the rand. We also conduct empirical tests for a cross-section of countries. We expect to observe systematic differences in the spreads across currencies as country risk varies. Section 6 summarizes and concludes our findings.
2. Bid-ask spreads and exchange rate volatility
We know that the time series process of high-frequency exchange rate data exhibits clusters of volatility. Such a finding is consistent with the results of previous time series studies of exchange rates such as those reviewed by Bollerslev et al. (1992). The autocorrelation of volatility means that as the conditional variance of the exchange rate increases today, dealers forecast a greater variance in the future. Following the modeling strategy of Glosten and Milgrom (1985), Admati and Pfleiderer (1989), and Andersen (1993), Bollerslev and Melvin (1994) present a simple theoretical model of the relationship between volatility and the bid-ask spread. The model is briefly summarized here to motivate the empirical work below.
Assume that the foreign exchange market is made up of liquidity traders, whose transactions are driven by the needs of financing trade, and informed traders, who possess superior information either from information advantages gained from transactions with liquidity traders or information asymmetries regarding the relevant fundamentals. The fundamental value of the spot rate is assumed to evolve over time as a martingale, but the informed traders know next period's shock to the spot rate. At time t-1, traders will set bid and ask quotes for transactions at time t (the market maker is obliged to trade up to some maximum amount of currency at existing quotes). The existing quotes will generate losses for the market maker when the counter-party is an informed trader. Trades with liquidity traders will earn the spread. Standardizing the innovations to the spot rate introduces the conditional variance into the model and the zero profit condition written in terms of the spread has the spread changing in proportion to the volatility of the underlying spot rate.
Since the bid-ask spread is priced based on uncertainty regarding future movements in the spot rate, a greater expected conditional variance of the spot rate should result in a larger spread on current quotes. This suggests that the appropriate empirical modelling strategy is a GARCH framework to take into account the autocorrelation of the conditional variance.
Before proceeding to empirical estimation, we should examine the raw data for evidence of covariance stationarity to guide our choice of functional form. For our time-series estimation we will focus on the South African rand as a good example of a currency that has been subjected to repeated large shocks due to sociopolitical unrest. We will later turn to cross-section evidence for a larger group of currencies to test for political risk effects on bid-ask spreads.
Daily bid and ask prices of the South African rand against the US dollar were obtained from the London Financial Times over the period from October 2, 1985, to August 31, 1990, for a total of 1,242 observations. We extract daily spot rates as the mean of the bid and ask prices. Table 1 presents some descriptive statistics on the exchange rates and spreads. Note that the mean daily changes in spot rates do not differ significantly from zero. However, the spreads differ significantly from zero. The largest spread observed for the rand is almost 9%.
The skewness (which would equal zero for a normal distribution), kurtosis (equal to 3 for a normal distribution), and Bera-Jarque statistics all indicate that we can reject the hypothesis of normally distributed data. The Ljung-Box Q tests (appropriate for the non-normally distributed data explored here) indicate strong evidence of autocorrelation in the spread data. The Ljung-Box [Q.sup.2] statistics also indicate autocorrelation in the second moments of all series. As well documented by many previous researchers, time-series estimates of high-frequency exchange-rate models should account for the autoregressive heteroscedasticity that is generally present.
Mean 0.0235 2.7832 (0.42) (95.26) Standard deviation 1.9998 1.0292 Skewness -0.16 0.81 Kurtosis 7.75 3.63 B-J(***) 1,172.90 168.41 Q(12)([section]) 12.71 10,009.57 [Q.sup.2](12)([section]) 137.97 9,063.27
* Daily percentage changes in the rand/dollar exchange rate.
** [100.sup.*] [Ask - Bid]/[Ask + Bid]/2].
*** B-J is the Bera-Jarque test for normality. It is distributed [Mathematical Expression Omitted] under the null hypothesis of normality, and is computed as
T[[skewness.sup.2]/6 + [(kurtosis - 3).sup.2]/24]
The 95% critical value is 5.99.
[section] Ljung-Box Q test for autocorrelation. The 95% critical value of 12 d.f. is 21.03.
t-Statistics are in parentheses.
Time-series model of the rand
To accommodate the observed serial correlation in both the conditional first-order and second-order moments, a multivariate MA(1)-GARCH(1, 1)-M process is estimated for the relationship between the spreads and the time-varying conditional variance. Specifically, the model to be estimated is
[Mathematical Expression Omitted]
[Mathematical Expression Omitted]
[Mathematical Expression Omitted]
[Mathematical Expression Omitted]
where [s.sub.t] and [v.sub.t] represent the log-differenced spot rate and the percentage spread and [Mathematical Expression Omitted] and [Mathematical Expression Omitted] represent the conditional variance of the spot rate and spread, respectively and C[R.sub.t] represents specific proxies for country risk to be discussed below.(1)
The model is first estimated without explicit country risk proxies to infer the effect of spot rate volatility, as measured by the conditional variance, on the spread. Estimation results are given in the first column of Table 2. The [Mathematical Expression Omitted] coefficient in the [Mathematical Expression Omitted] equation indicates the significance of the conditional variance of the spot rate in forecasting the future conditional variance, as hypothesized above. In the estimation of the spread as a function of the conditional variance, the [Mathematical Expression Omitted] term in the [v.sub.t] equation implies that the spread on the rand widens with greater conditional variance of the current period spot rate. The Q(12) and [Q.sup.2](12) test statistics imply that all residuals have been reduced to white noise and these, along with the LR test statistics, indicate that the model can be considered well-approximated by the GARCH process.
The results, so far, have supported the hypothesis of autocorrelated volatility and the positive effect of volatility on the spread. However, we have yet to explicitly incorporate social and political events into the analysis. We hypothesize that the volatility witnessed in this market is frequently a result of socio-political crises in South Africa. A formidable task is to specify meaningful measures of socio-political turmoil that would likely cause volatility to shift in the manner described.
[TABULAR DATA FOR TABLE 2 OMITTED]
In the case of South Africa, we focus on domestic turmoil that has largely resulted from racial policies. Data were gathered on a daily basis for four variables:
(i) Riots: violent demonstrations or clashes involving a large group;
(ii) Antigovernment demonstrations: organized, nonviolent gatherings of large groups for the purpose of protesting against government actions or leaders;
(iii) Political strikes: strikes by workers or students for the purpose of protesting against the government.
(iv) Armed attacks: acts of violence using weapons that are aimed as protests or rebellion against the government.
We expect a fairly high degree of collinearity among these events. As a result, we construct a dummy variable (CR) equal to 1 when one of the events occurs, and equal to zero otherwise.
We may think of these variables as country risk measures, as they record events that change the perception of the risk associated with financial transactions involving the countries. Country risk may be defined as the risk of an unexpected change due to political or social conditions that adversely affects the returns to foreign investors. Daily data on the country risk variables are collected from the New York Times Index and the London Times Index. Out of approximately 1200 observations for South Africa, there were 201 days with riots, 50 days with demonstrations, 30 days with political strikes, and 82 days with armed attacks.
One potential problem with the measures of country risk is that no adjustment is made for the fact that some events will be more important and have greater effects than others. We try to adjust for this differing intensity or importance of events for South Africa by interacting the country risk dummy variable with the number of deaths reported as resulting from the events. During the period studied, there were 980 deaths reported to result from attacks. We report empirical results below for both the unadjusted country risk proxy (CR) and risk adjusted by the number of associated deaths (CR*Deaths).
We consider two alternative ways of incorporating the country risk variables into our model. First, to test our hypothesis that volatility in the foreign exchange value of the rand is associated with sociopolitical crisis, we directly include the variables in the conditional variance equation of the spot exchange rate. Columns 2 and 3 of Table 2 report the estimation results. In this case, we test the hypothesis that the country risk variables are used to forecast the variance and are, thus, determinants of the conditional variance. In both cases, the coefficients on CR and (CR*Deaths) are statistically significant and positive. The evidence indicates that the greater the intensity of country risk events, the greater the conditional variance of spot exchange rates, and the greater the associated foreign exchange market bid-ask spread.
Second, we examine the hypothesis that the country risk variables may affect the spreads directly, independent of the volatility effect. We test if there is any independent influence of country risk on the spread beyond that embodied in the conditional variance by including the CR and (CR*Deaths) variables directly in the conditional mean spread equation. This may capture the effect of news-driven spreads apart from the volatility-induced spreads. The estimated equations are reported in columns 4 and 5 of Table 2. Note that while the country risk terms are still highly statistically significant in the conditional variance equation of the spot exchange rate, they are also significant (at lower levels of significance) in the conditional mean equation of the bid-ask spread. It appears that there do exist country risk news effects in addition to pricing the effects of volatility. The positive signs of CR and (CR*Deaths) on [v.sub.t] in columns 4 and 5 in Table 2 indicate that the news effect is to increase the spread on the rand. In addition to this direct news effect on the mean, there is also the effect of the news on uncertainty as proxied by the conditional variance.
The estimated coefficients in Table 2 indicate that, other things being equal, on days where country risk events occur, the spread will average 0.1999 rand higher, or 540 basis points higher, than on tranquil days. This is almost twice the mean spread of 0.1029, so that the spread is almost three times the 'normal' spread on days associated with a country risk event. In addition, an extra death due to politically-motivated violence will cause the spread to rise by 0.1481 rand, or 400 basis points.(2)
4. Cross-section evidence for many currencies
If foreign exchange market spreads are priced in the manner described in Section 3, then we should observe systematic differences in the spreads across currencies in addition to the differences across time that were just documented. To investigate the evidence across currencies, we obtained daily bid and ask prices for 36 currencies over the period of March 1987 to August 1990 from Data Resources, Inc. The 36 countries chosen have country risk indexes that evaluated the level of risk associated with each country as published in the International Country Risk Guide (ICRG). Since the country risk indexes are only updated monthly, we construct monthly averages of the daily percentage bid-ask spreads along with the monthly standard deviations of the daily percentage changes of the spot exchange rates for all 36 currencies.
Table 3 describes how the country risk indexes are constructed. Thirteen categories are evaluated and a score awarded based on the perceived position of each country. The higher the score, the better conditions in the country. Lower scores, then, indicate greater country risk.
Table 4 lists the countries for which data on spreads and country risk are available. Note that there is a mix of developing and industrial countries. The second column lists the average daily percentage spread over the March 1987-August 1990 period. The third column lists the average monthly standard deviations of daily percentage changes in the exchange rates. The fourth column lists the average total value of the country risk index for each country. The lower the value of the index, the greater the country risk. Finally, the fifth column provides a ranking of countries based on their average country risk index values. According to the ICRG, during this period, Israel was the riskiest of the 36 countries studied here (although not the riskiest country covered by the ICRG. Switzerland was the least risky country.
To estimate the effect of the country risk indexes on spreads, we first ran a cross-sectional regression for each month in the sample across the 36 country observations. The independent variables included a constant, the standard deviation of the exchange rate changes, and the country risk indexes. This preliminary set of regressions indicated that the effect of country risk seemed to vary over time. Particularly, the coefficients on the country risk indexes were larger for more recent months than earlier months. We tested the hypothesis that the relationship between the spreads and their determinants experienced a structural change by including dummy variables equal to zero prior to June 1989 and then one afterward. June 1989 is chosen as a period of structural break as the Tiananmen Square conflict in China was associated with large capital flows and much official intervention in the foreign exchange market. The significance of the dummy variable in this preliminary exercise indicates that the Tiananmen Square conflict is associated with a significant structural break in the relationship between bid ask spreads and country risk.
Description of various risk indices
Index Description Maximum points
A Economic expectation versus realities (12) B Economic planning failure (12) C Political leadership (12) D External conflict risk (10) E Corruption in government (6) F Military in politics (6) G Organized religion in politics (6) H Law and order tradition (6) I Racial and nationality tensions (6) J Political terrorism (6) K Civil war risks (6) L Political party development (6) M Quality of the bureaucracy (6)
Source: International Country Risk Guide (ICRG).
Total points for each of the political risk indicators out of the maximum points are indicated in parentheses.
Identifying June 1989 as a period of structural break results in our combining the time series and cross-section data and estimating a panel data model for two periods, March 1987-May 1989 and June 1989-August 1990. Specifically, a random-effects model was estimated for the following equation
[v.sub.it] = [[Alpha].sub.0] + [[Alpha].sub.1][[Sigma].sub.it] + [[Alpha].sub.2]C[R.sub.it] + [[Epsilon].sub.it] + [u.sub.i] (5)
where C[R.sub.it] represents the value of the particular measure of country risk for country i in month t, [[Sigma].sub.it] represents the standard deviation of daily changes in exchange rates for each country i in month t, and [u.sub.i] is a country specific disturbance which could be viewed as the collection of factors not in the regression that are specific to that country. There are 13 categories on which the country risk indexes are based. The estimation results are reported in Table 5. Due to the collinearity among the different categories of country risk, we regress the spreads on each individual category as well as on the total index [TABULAR DATA FOR TABLE 4 OMITTED] value and then report regression results for all categories in one equation at the bottom of the table.
Table 5a reports the results over the March 1987-May 1989 period and Table 5b contains the results for the June 1989-August 1990 period. Note that the standard deviations of the spot exchange rates are all statistically significant. [TABULAR DATA FOR TABLE 5A OMITTED] [TABULAR DATA FOR TABLE 5B OMITTED] The greater the standard deviation of the percentage changes in the exchange rate, the greater the foreign exchange market bid-ask spread across countries. The individual components of the country risk index are all statistically significant as is the total index in Table 5a.
Including all of the components in one regression illustrates the effect of multicollinearity - the data are weak in assigning individual effects when all are included. In Table 5b, only the index I, 'Racial and Nationality Tensions' is individually insignificant. All other components, along with the total index, indicate that the lower the country's rating, the greater the spread. So countries with greater risk will have currencies that are traded with larger spreads, as predicted by the theory of Sections 2 and 3. The effects of the various significant components range from a six basis point increase in the percentage spread due to a one point drop in the index for political party development in Table 5a, to a 120 basis point increase in the percentage spread due to a one point drop in the index for quality of the bureaucracy in Table 5b. The total index in Table 5a indicates that there will be a 2.9 basis point increase in the percentage spread for each one point fall in the index. The overall index in Table 5b indicates that there will be a 7.4 basis point increase in the percentage spread for each one point fall in the index.
We have provided evidence that posted spreads in the foreign exchange market do respond to changes in market conditions. The specified market conditions that we address are changes in country risk due to social or political unrest. Our hypothesis is that sociopolitical unrest contributes to greater uncertainty regarding future exchange rates so that dealers will increase the bid-ask spreads offered.
The empirical evidence examines both time series changes in the spread for the South African rand as well as cross-section differences in the spread across a mix of industrial and developing country currencies. The time-series evidence indicates that country risk variables like riots, demonstrations, strikes, armed attacks, and related deaths lead to an increase in the spread directly as well as indirectly through an increase in the conditional variance of exchange rates which, in turn, causes the spread to widen. The cross-section evidence indicates that the spread is greater for currencies of countries with a greater perceived amount of country risk (the risk of an unexpected change in the social or political situation that would adversely affect the return to foreign investors). So both the changes in the spread over time for particular countries and the changes in the spread across countries at a particular time appear to be significantly related to country risk differences.
What is the price of social or political unrest? The time-series evidence suggests that the spread on the South African rand tends to rise by 540 basis points on days when South Africa experiences riots, demonstrations, strikes, or armed attacks. The spread on the rand also tends to rise by 400 basis points with a death due to politically-motivated violence. Across countries, the price of country risk is estimated to range from a 6 to 120 basis point increase, with changes in the quality of the bureaucracy having the largest effect in increasing a currency's spread.
1 Ideally, one would like to be able to estimate a version of the model where the lagged squared residuals and conditional variances of both variables are included in eqs (3) and (4), along with an additional equation for the conditional covariance. Unfortunately, such a model specification results in the variance-covariance block of the model having 21 parameters to estimate. We were unable to achieve convergence with a model this complex. To consider the potential relationship between [Mathematical Expression Omitted] and [Mathematical Expression Omitted], we estimated a version of the model including a conditional covariance equation in the system with lagged values of the conditional covariance included. The covariance equation added nothing of statistical significance to the model and the qualitative results of the more parsimonious representation reported in Table 2 were unchanged.
2 Since the equations are estimated using the percentage spread, the effects on the level of the spread are evaluated using the mean values. For instance, using column 4 of Table 2, dv/dCR is equal to (dv/dCR) directly in the mean equation plus the volatility effect of [Mathematical Expression Omitted] for a total effect of 0.0540 for CR unadjusted for deaths. If the mean v equals 0.0278, then an increase of 0.0540 (540 basis points) raises v to 0.0818. With a mean mid-price of 3.7021 rand per dollar, we infer that the spread rises from 0.1029 rand to 0.3028, a rise of 0.1999. Similarly, we find that the effect of an extra death raises the percentage spread 400 basis points or 0.1481 rand.
We acknowledge the helpful comments of Catherine Carey, Bettina Peiers, colleagues at Northwestern University (where Melvin was a Visiting Professor of Finance when this work began), two anonymous referees, and participants at the Western Economic Association session on 'Bid-Ask Spreads in the FX Market' in Seattle. We thank Jahangir Sultan, for providing a multivariate GARCH program along with advice on its use, and Bettina Peiers, for her excellent research assistance.
ADMATI, A. R. and PFLEIDERER P. (1989). 'Divide and Conquer: A Theory of Intra-day and Day of the Week Effects', Review of Financial Studies, 2, 189-223.
ANDERSON T. G. (1993). 'Return Volatility and Trading Volume: An Information Flow Interpretation of Stochastic Volatility' Working Paper, Kellogg Graduate School of Management, Northwestern University, Evanston, IL.
BOLLERSLEV T., CHOU, R. Y., and KRONER, K. F. (1992). 'ARCH Modeling in Finance: A Review of the Theory and Empirical Evidence', Journal of Econometrics, 52, 5-59.
BOLLERSLEV, T. and MELVIN, M. (1994). 'Bid-Ask Spreads and Volatility in the Foreign Exchange Market', Journal of International Economics, 36, 355-72.
BOOTHE, P. (1988). 'Exchange Rate Risk and the Bid-Ask Spread: A Seven Country Comparison', Economc Inquiry, 26, 485-92.
GLASSMAN, D. (1987). 'Exchange Rate Risk and Transactions Costs: Evidence from Bid-Ask Spreads', Journal of International Money and Finance, 6, 479-90.
GLOSTEN, L. and MILGROM P. (1985). 'Bid, Ask, and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders', Journal of Financial Economics, 13, 71-100.
GOODHART, C. A. E. and FIGLIUOLI, L. (1991). 'Every Minute Counts in Financial Markets', Journal of International Money and Finance, 10, 23-52.
International Country Risk Guide, The Economist Intelligence Unit, various monthly issues.
|Printer friendly Cite/link Email Feedback|
|Author:||Melvin, Michael; Tan, Kok-Hui|
|Publication:||Oxford Economic Papers|
|Date:||Apr 1, 1996|
|Previous Article:||Economic growth, convergence clubs, and the role of financial development.|
|Next Article:||Ownership and efficiency in nuclear power production.|