Conditional and unconditional intraday value-at-risk models: an application to high-frequency tick-by-tick exchange-traded fund data

Given the growing popularity of exchange-traded funds (ETFs) in the trading environment, this paper evaluates the performance of unconditional and conditional intraday value-at-risk (IVaR) models for high-frequency tick-by-tick ETF data. Three skewed distributions (the skewed t distribution, the skewed generalized error distribution and the skewed generalized t distribution) are used to generate unconditional IVaR. For conditional IVaR, a hybrid method combining the autoregressive conditional duration (ACD) model and the generalized autoregressive conditional heteroscedasticity (GARCH) model is considered. We analyze 36 variations of the ACD-GARCH model, 35 of which have not been investigated before. Backtesting results show that the unconditional models are appropriate for the first 30 minutes of the trading session, emphasizing the need for a time-varying mean and volatility at longer forecast horizons. The conditional ACD-GARCH models underestimate the risk for midcap ETFs but are suitable for one-day ahead forecasting for small- and large-cap ETFs. The choice of the ACD form (ie, linear or nonlinear) and the error distribution have no impact on the validity of the ACD-GARCH model. Our results are of interest to financial practitioners involved in high-frequency trading and who rely on intraday risk measures to maintain sufficient capital to absorb and mitigate potential losses.