An empirical investigation of U.S. stock prices : jumps and co-movements

• Main Author: Gilder, Dudley
• Published: Lancaster University 2011
• Subjects: 332.63
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.652026

The introduction of high-frequency data has facilitated the development of new financial econometric techniques, especially in nonparametric volatility measurement through the construction of realised volatility (RV). New non parametric techniques have also becn created for detecting jumps. In t his thesis the application of nonparametric jump tests to individual stock price series and the existence of cojumps, simultaneous jumps in the prices of more than one asset, is explored. Also, the relative merits of high-frequency data in forecasting the covariance matrix are examined. This research is contained in three chapters whilst another two comprise literature reviews on the econometric detection of jumps and covariance modelling. Chapter 3 develops a method for cojump detection based on the application of the intraday jump test of Andersen, Bollerslev and Dobrev (2007) to univariate price series. A coexceedance criterion, where a jump is identified for the same intraday interval across more than one asset, is then used to detect the presence of a cojump event. The presence of an intraday volatility pattern is also corrected for explicitly, allowing the jump tests to be applied at a higher significance level. One of the main advantages of this approach is that the stocks involved in each cojump event can be identified explicitly. This allowed systematic cojumps, idiosyncratic cojumps and singular jumps to be identified. The relative frequency of each of these types of cojump and jump was then translated into the relative importance of market, industry and firm-specific news in generating jumps. The intraday nature of the jump tests also meant three hypotheses based on the extant literature could be evaluated: Hypothesis 1 stated that jumps in the market portfolio and systematic cojumps were rare, Hypothesis 2 stated that jumps and cojumps would be negative on average and Hypothesis 3 stated that more jumps and cojumps would cluster around 10:00:00. Chapter 4 is a simulation study to determine the impact of employing different spot volatility estimators in the intraday jump tests of Andersen, Bollerslev and Dobrev (2007) (ABD) and Lee and Mykland (2008) (LM). Three alternative estimators were employed: A block sampling estimator (BS) based on the approach of ABD, a rolling window estimator (RW) based on the approach of LM and a GARCH estimator. A total of 10 different specifications for the true price process were used, where the sizes of both price and volatility jumps were altered between the specifications. To make the simulated prices reflect empirical reality, an intraday volatility pattern and microstructure noise were introduced. This enabled the effects of correcting the intraday volatility pattern and of sparse sampling to mitigate the effect of microstructure noise were also investigated. Lastly, Chapter 6 investigates the relative performance of models based on highfrequency returns, daily returns and option data in forecasting the covariance matrix. Comparisons of 7 models were made. Two of these were benchmark models based on the sample covariance matrix estimated using high-frequency and daily returns. To be able to use option data, returns were assumed to be generated by a market model. A forecast of the covariance matrix was then constructed from risk-neutral betas (Chang. Christoffersen, Jacobs and Vainberg, 2009) and model-free volatility (Bakshi, Kapadia and Madan, 2003) (option-factor-implied model). The remaining four models were dynamic models, two based on high-frequency returns and two based on daily returns. One high-frequency (Factor ARMA) and one daily return (Factor Double ARCH) based model also assumed returns were generated by the market model so that any detrimental effect arising from making this assumption could be isolated. The final high-frequency based model used the Cholesky decomposition to ensure forecasts were positive definite (CF-ARMA model) and the finaJ daily return based model was the popular, multivariate GARCH, DCC model of Engle (2002). Since the loss function of the agent using the forecasts is unknown, three evaluation criteria were used to assess forecasting performance. The first was the MSE (statistical loss functon), the second was Mincer-Zarnowitz regressions and third was the variance of hedging portfolios constructed using the covariance matrix forecasts (economic loss function).