APPENDIX I Overview of Political and Economic Developments
This paper was prepared while Christian Jochum was visiting the IMF. He is a Ph.D. student at the University of St. Gallen and holds an M.Sc. from the University of Warwick. Laura Kodres holds a Ph.D. from Northwestern University and is an Economist in the IMF’s Research Department. The authors are grateful to Carlo Cottarelli, Robert Flood, Graeme Justice, Charles Kramer, Ricky Lam, David Ng, Lorenzo Perez, Patricia Reynolds, S. Hossein Samiei, Garry Schinasi, Gabriel Sensenbrenner, and Philip Young for helpful comments and discussion during the preparation of the paper, to Darvas Zsolt, of the Hungarian National Bank for the provision of the Hungarian data, and to the Malaysian authorities for the initial prompt to undertake the study. They are further grateful to James Hamilton.
Other related elements often identified as providing low transaction costs include lower fees and commissions, lower opportunity cost of initial margin, lower opportunity cost of additional liquid assets held to meet variation margin, smaller bid-ask spreads, and fewer regulatory constraints.
While several other emerging market currencies have recently begun trading (or are scheduled to), their price history is still too short to use in any empirical work.
For the U.S. market conflicting evidence is presented on the question of whether the options market leads the stock market. Stephan and Whaley (1990) present results favoring this conclusion, while Finucane (1991) reaches the opposite conclusion.
McCarthy and Najand (1993) find a positive relation between futures trading volume and futures volatility.
Kawaller, Koch, and Koch (1993) also report a strengthened relation between cash and futures markets as the futures market volatility rises. They attribute this behavior to the speed of information processing.
Jabbour(1994) states that the implied spot rates derived from futures prices are good predictors of future spot rates. This result can be regarded as circumstantial evidence that the derivative market is not excessively volatile.
The crash of the European Monetary System (EMS) in 1992 is a much-cited example for the destructive destabilization emanating from derivative markets. Nevertheless, it could be argued that futures and options markets were anticipating and thus accelerating realignments, which had been postponed for too long.
The tendency of ARCH models to imply too much volatility persistence was demonstrated in the analysis of the October 1987 stock market crash, for example, Engle and Mustafa (1992).
It may be possible to make these transition probabilities a function of macro-economic variables associated with regime shifts, thereby linking this purely time-series model of volatility to economic fundamentals. However, since our intention is simply to obtain a base case against which to measure the effect of a futures introduction, this extension is left for future research.
The response of an element i in zt to a shock in another element j in zt, is described by the sequence 𝛙i, j,1,𝛙i, j,2, M𝛙i, j,3,….the impulse response function.
Many emerging market countries, including those examined here, manage their currencies so as to remain within a band. The use of the SWARCH methodology would, if the band effectively limited ex post volatility to be within a constant range over the sample period, show that only one “regime” would be necessary to accommodate the time-series pattern of volatility. In this case, the SWARCH model is superfluous and the use of an ARCH model would suffice. Thus, since the SWARCH model is purely a statistical model for volatility, the existence or nonexistence of a formal (or informal) exchange rate band does not affect its usefulness. The data determine whether multiple regimes are needed to provide a good statistical fit.
The data were provided by Bloomberg, the Futures Industry Institute, and the Hungarian National Bank.
Daily open interest data are not available on the Hungarian forint/U.S. dollar contract.
Because the nearby series will have discrete jumps at transition points between contract months, dummy variables for these dates are introduced in the return equation. The results are not sensitive to this method for estimating the return equation.
Futures data on the Mexican peso are only available after April 1995, since the contract began trading at this time.
The sample lengths are relatively short for unit root tests to have high power. However, despite the added tendency to accept the null hypothesis of a unit root when it is not present, we strongly reject the presence of a unit root.
Only the futures return series for the Hungarian forint does not show the regime switching characteristics initially assumed and thus the SWARCH model is not used in this case. A three-regime model was attempted, but did not converge for any of the series.
Using a shorter lag structure results in very slow convergence and parameter estimates were not robust to alternative starting values. Lag lengths between two and five lags have been investigated.
Note that this feature of the spot and futures prices–that the spot prices remain unchanged for several days at a time while the futures prices are rarely unchanged–by itself suggests that futures markets incorporate information faster than spot markets.
Since the appropriate length of time over which agents measure and react to volatility may be different from the daily horizon assumed here, the results for the Mexican peso have been reestimated using returns measured over five trading days. The smaller number of observations typically lowers the significance of the estimated coefficients and the model utilizes a normal distribution, rather than the fatter-tailed t distribution, to obtain convergence. However, the results are qualitatively the same: ARCH effects are still present; there are significant regime shifts; and similar probabilities for the transition matrices are obtained. This result accords with other studies–Droste and Nijman (1993) and Diebold (1988)–that show that ARCH effects are relatively stable at multiple sampling frequencies.
There were a number of large outliers in the data. To gauge their influence on the results, the model is reestimated for the Mexican peso spot rate after removing the four largest outliers (in absolute value) and replacing them with the average value of the previous and following observation. The values of P1 and P2 are virtually unchanged and the correlation coefficient between the estimated volatility of the series with and without the outliers is 0.92. Only the second ARCH coefficient changes appreciably, from 0.203 to 0.112, while the other estimates remain more or less unchanged. From these results, we view the influence of outliers as limited.
The optimization procedure was run in GAUSS 3.0 employing the OPTMUM package. Some of the routines are derived from programs generously provided by J. Hamilton.
Since exchange rates are the price of one currency in terms of the other and can be quoted in dollar terms, or the reciprocal, a leverage effect is thought to be less likely for exchange rate series. However, one could argue that, for emerging market currencies, where the numeraire currency is the dollar, the distinction between a depreciation and an appreciation in local currency may be meaningful. The absence of a leverage term demonstrating this effect may be a reflection of the time period used– all three currencies were depreciating and were expected to maintain this trend.
The use of “generated” variables in subsequent econometric techniques can sometimes be problematic. We believe we have circumvented any biases due to the generated volatility estimates by not using any common variables from the original spot and futures SWARCH specifications in subsequent specifications.
While we attempt to use sound economic arguments to appropriately choose the ordering of the variables in the VAR, it is useful to note that the covariance terms in the VAR error variance-covariance matrix are typically quite low, ranging from 0.009 to 0.47 with a mean of 0.12. Since these coefficients are generally low, the orthogo-nalizauon and, consequently, the exact methodology used are of somewhat reduced importance. This is borne out by reestimating the variance decomposition using the Bemanke (1986) approach in which the ordering of the variables is replaced by an ordering of the errors, thereby generating an impulse response function that is not influenced by the original ordering of the variables. The results are substantively the same. Further, reestimating the variance decomposition with the positions of spot volatility and futures volatility exchanged does not significantly alter the results in Table 6: for all three currencies and for any ordering chosen over the 10-day horizon, the own-variance of the spot market variable is never lower than 60 percent. Moreover, the futures market variables continue to have strong self-explanatory power as well.
With the minor exception of spot on spot in the Mexican peso case, which most probably is due to the high starting level of 98.23 percent.
Our futures data run from January 1, 1995 through February 28, 1997, limiting the sample period for the previous results. However, our spot series is longer, allowing us to expand the sample for this purpose.
The use of dummy variables in (G)ARCH type volatility estimation was first outlined by Baillie and Bollerslev (1989), who use dummies to account for day of the week effects. The dummy d, takes the value of 1 with the introduction of futures contracts.
Replacing the simple step dummy with a measure of futures market volume decreases the coefficient in absolute size, but does not affect the level of significance or the sign of the coefficient.
Replacing the volume variable in models (5a) and (5b) with the SWARCH estimate for futures market volatility again yields negative coefficients on the futures variable, which lack statistical significance in two of the three cases. However, these results were numerically difficult to obtain and appear to lack the robustness of our other results