Estimation and forecast evaluation of risk measures with high frequency financial data

This thesis contributes to the financial econometric literature in the areas of estimation, forecasting, and forecast evaluation, using high frequency financial data. The thesis focuses on the use of this data to estimate risk parameters commonly described at lower frequencies, e.g. using intraday data to estimate daily variance or daily value-at-risk. A common theme in all chapters is that the use of high frequency data can dramatically improve the solutions to common financial problems. Chapter 2 demonstrates how a dependent bootstrap can be used to consistently estimate a wide range of risk measures associated with a daily return, given a sequence of intraday returns. Estimable parameters include variance, value-at-risk, expected shortfall, semi-variance, skewness, kurtosis, and robust risk measures. Excluding the case of variance, all estimators are, to the best of my knowledge, the first of their kind in the literature: non-parametric, and consistent, in the presence of market microstructure noise. The theory also contains a new result on the convergence of bootstrapped parameters that is more generally applicable in the theoretical literature on dependent bootstraps. Chapter 2 also demonstrates an application of the proposed estimation methodology. The method is used to construct a consistent proxy for value-at-risk which is used to rank value-at-risk forecast models in a dual-asymptotic framework. The approach is shown via both simulation and empirical work to exhibit much greater power to distinguish between competing value-at-risk forecasts than other tests in the literature. Chapter 3 extends the empirical application in Chapter 2 to 351 value-at-risk forecast models, and to a larger dataset which spans two exchanges and two distinct forecasting intervals. A new class of value-at-risk forecast models based on the estimation methodology from Chapter 2 are proposed, and are shown to provide more accurate forecasts than all other models under consideration. More generally, the results strongly suggest that value-at-risk forecasts that utilise simple time series models of proxies based on intraday data significantly outperform forecasts which utilise daily data exclusively. Chapter 4 proposes a data-based method for ranking variance estimators constructed from intraday data. This paper draws from the literature on loss-based forecast evaluation, but accounts for the inevitable dependencies that occur when ranking estimators as opposed to forecasts. Under certain conditions, the method is shown via simulation to exhibit greater power than other methods in the literature. The chapter also contains a new technical result on the product of near epoch dependent processes that is widely applicable in the time-series literature.