I perform comprehensive comparison of the standard realised volatility estimators including a novel wavelet time-frequency estimator (Barunik and Vacha 2012) on wide variety of assets: crude oil, gold and S&P 500. The wavelet estimator allows to decompose the realised volatility into several investment horizons which is hypothesised in the literature to bring more information about the volatility time series. Moreover, I propose artificial neural networks (ANN) as a tool for forecasting of the realised volatility. Multi-layer perceptron and recursive neural networks typologies are used in the estimation. I forecast cumulative realised volatility on 1 day, 5 days, 10 days and 20 days ahead horizons. The forecasts from neural networks are benchmarked to a standard autoregressive fractionally integrated moving averages (ARFIMA) model and a mundane model. I confirm favourable features of the novel wavelet realised volatility estimator on crude oil and gold, and reject them in case of S&P 500. Possible explanation is an absence of jumps in this asset and hence over-adjustment of data for jumps by the estimator. In forecasting, the ANN models outperform the ARFIMA in terms of information content about dynamic structure of the time series.

The focus of this thesis is on the employment of theoretical and practical quantile methods in addressing prediction, risk measurement and inference problems. From a prediction perspective, a problem of creating model-free prediction intervals for a future unobserved value of a random variable drawn from a sample distribution is considered. With the objective of reducing prediction coverage error, two common distribution transformation methods based on the normal and exponential distributions are presented and they are theoretically demonstrated to attain exact and error-free prediction intervals respectively. The second problem studied is that of estimation of expected shortfall via kernel smoothing. The goal here is to introduce methods that will reduce the estimation bias of expected shortfall. To this end, several one-step bias correction expected shortfall estimators are presented and investigated via simulation studies and compared with one-step estimators. The third problem is that of constructing simultaneous confidence bands for quantile regression functions when the predictor variables are constrained within a region is considered. In this context, a method is introduced that makes use of the asymmetric Laplace errors in conjunction with a simulation based algorithm to create confidence bands for quantile and interquantile regression functions. Furthermore, the simulation approach is extended to an ordinary least square framework to build simultaneous bands for quantiles functions of the classical regression model when the model errors are normally distributed and when this assumption is not fulfilled. Finally, attention is directed towards the construction of prediction intervals for realised volatility exploiting an alternative volatility estimator based on the difference of two extreme quantiles. The proposed approach makes use of AR-GARCH procedure in order to model time series of intraday quantiles and forecast intraday returns predictive distribution. Moreover, two simple adaptations of an existing model are also presented.

Realised Stochastic Volatility model of Koopman and Scharth (2011) is applied to the five stocks listed on NYSE in this thesis. Aim of this thesis is to investigate the effect of speeding up the trade data processing by skipping the cleaning rule requiring the quote data. The framework of the Realised Stochastic Volatility model allows the realised measures to be biased estimates of the integrated volatility, which further supports this approach. The number of errors in recorded trades has decreased significantly during the past years. Different sample lengths were used to construct one day-ahead forecasts of realised measures to examine the forecast precision sensitivity to the rolling window length. Use of the longest window length does not lead to the lowest mean square error. The dominance of the Realised Stochastic Volatility model in terms of the lowest mean square errors of one day-ahead out-of-sample forecasts has been confirmed.

The problem of determining whether or not a theoretical model is an accurate representation of an empirically observed phenomenon is one of the most challenging in the empirical scientific investigation. The following study explores the problem of stochastic model validation. Special attention is devoted to the unusual two-peaked shape of the empirically observed distributions of the conditional on realised volatility financial returns. The application of statistical hypothesis testing and simulation techniques leads to the conclusion that the conditional on realised volatility returns are distributed with a specific previously undocumented distribution. The probability density that represents this distribution is derived, characterised and applied for validation of the financial model.

Cette thèse est consacrée à l’étude de théorèmes limites : grandes déviations et déviations modérées pour des estimateurs liés à des modèles financiers. Dans une première partie, nous nous sommes intéressés à l’étude des déviations grandes et modérées des estimateurs de la covariation et de la (co)volatilité réalisée issus des fonctionnelles associées à deux processus de diffusion couplés de manière synchronisée. Les techniques utilisées dans ces travaux sont basées d’une part sur celles utilisées dans Djellout-Guillin-Wu et sur la sous additivité et sur la notion d’approximation exponentielle inspirées des travaux de J. Najim d’autre part. Dans une deuxième partie, on considère que les deux processus de diffusion sont observés de manière non synchronisée et on établit des déviations modérées pour l’estimateur de la variation généralisée et pour celui de Hayashi-Yoshida. Les résultats sont obtenus par l’utilisation d’une nouvelle approche sur les déviations modérées des variables aléatoires m−dépendantes vérifiant des conditions de type "Chen-Ledoux". Dans la troisième et dernière partie, on s’intéresse à l’étude processus autorégressif d’ordre p dont le bruit est un processus autorégressif d’ordre q. On montre des déviations modérées pour certains estimateurs associés à notre modèle dont la statistique de Durbin-Watson. Les résultats sont donnés dans le cas où le bruit est gaussien puis dans le cas de condition de type "Chen-Ledoux" portant sur le bruit. / This thesis is devoted to the study of the limits theorem : large and moderate déviations for some financial mathematicals estimators. In the first part, we studied the large and moderate deviations of the estimators of covariation and the realized (co)volatility obtained from the functional associated to two diffusion processes coupled in synchronous manner. The techniques used in this work are based, on the one hand, on those used in Djellout-Guillin-Wu and the subadditivity and the exponential approximation notion inspired by J. Najim results on the other hand. In the second part, we consider that ours two diffusion processes are observed in a nonsynchronized manner and on the establish the moderate deviations for the generalised bipower variation estimator and the Hayashi-Yoshida estimator. The results are obtained by using a new approach on the moderate deviations of the m−dependent random variables based on the Chen-Ledoux type condition. In the third and last part, we study the stable autoregressive process of order p where the driven noise is also given by a q-order autoregressive process. We prove the moderate deviations for some estimators associated with our model such as the Durbin-Watson statistic. The results are given in the case where the driven noise is the normally distributed then in the case where the driven noise satisfy a Chen-Ledoux type condition.