DYNAMIC CONDITIONAL SCORE MODELS: FORECASTING VOLATILITY OF EXCHANGE RATES

Abstract

The standard GARCH models used for forecasting volatility have been critiqued for their limitations in accommodating the conditional heavy-tailed properties of returns data, producing forecasts overly sensitive to large disturbances. The DCS models have been advanced as an alternative in this regard, with the inclusion of the score of a specified distribution in the dynamic volatility specification mitigating this effect and producing more accurate volatility forecasts relative to the GARCH models.

This study leverages on this branch of the literature and examines whether differences in performances of the DCS and GARCH remain when a recursive scheme is employed. We find evidence from the AUD/USD and USD/JPY that the edge of the DCS models over the GARCH models diminishes when an expanding window is implemented. We attribute this to the congruent purpose of the recursive scheme to the score- in accounting for changes in volatility levels as new observations become available.