Seminonparametric Methods for Modeling Conditional Volatility of Exchange Rate

We employ a seminonparametric (SNP) methodology in characterizing the conditional density of the exchange rate changes. The model selection procedure based on the BIC is used by moving upward along an expansion path. We find the semiparametric AR(4)-GARCH(2,2) model for the KRW/USD returns and the semiparametric AR(1)-GARCH(2,2) model for the JPY/USD returns as the BIC preferred SNP models. Simulations from the BIC minimizing SNP models seem to appropriately mimic the actual data. The time dependent heterogeneity of the actual data is recognized by the simulations from the semiparametric AR-GARCH-type models and the nonlinear nonparametric ARGARCH- type models. We show that it is important to take departures from the Gaussianity of the data into account in specifying conditional heterogeneity of the exchange rate returns process. We also provide evidence on the benefits from using the SNP models in estimating the conditional density function via simulations.