Volatility in financial markets: long memory, asymmetry, forecasting

In this thesis, we investigate several aspects of asset price volatility dynamics in financial markets. In Chapter 1, we focus on the long memory property of financial volatility and study whether the long memory in the volatility is true (genuine) or spurious. We address the problem of a correct identification of a memory structure of financial volatility by considering it in the context of temporal aggregation. Firstly, we generalize the up-to-date theoretical knowledge about temporal aggregation in long memory processes to show that the long memory property of ARFIMA series is invariant to temporal aggregation. Secondly, we conduct a Monte Carlo simulation experiment and provide a regression analysis of the experiment results in order to validate the established theoretical implications for the semiparametric GPH method of ARFIMA estimation. Finally, we analyze empirically the long memory property of volatility of the GBP/USD foreign exchange rate returns at various time scales. We focus on different established volatility proxies (such as absolute returns, squared returns and realized volatility) and use the semiparametric ARFIMA estimation methods to investigate the long memory dynamics of the volatility at various levels of temporal aggregation. Based on the theoretical implications, we formally test the hypothesis of equivalence of the estimated long memory parameters at different time scales. We have found evidence that volatility of the returns is a true long memory process as it is characterized by the same fractional differencing parameter across the observed time scales. The evidence is particularly strong in case of realized volatility. In Chapter 2, we investigate the connection between the phenomenon of volatility asymmetry and the asymmetry of investor attention to good and bad market news. As an advanced and direct measure of investor attention, we utilize the Search Volume Index (SVI) provided by the Google Trends service. We use a long span of daily data for a range of international stock market indexes and employ a methodological framework of SVAR and ARDL models to study the direction and timing of the asymmetric effects. Our findings indicate that both volatility and investor attention are similarly asymmetric in their response to market news represented by index return: a negative return has a stronger impact on both volatility and investor attention than a positive return of the same absolute magnitude. We provide new evidence of positive relationship between volatility and investor attention and demonstrate that the magnitude of the impact of investor attention on volatility is stronger during periods of negative returns. We show that, in the established theoretical framework, retail investor attention can contribute to volatility asymmetry and create temporary asymmetric volatility fluctuations but is unlikely to be responsible for permanent shifts in market volatility. In Chapter 3, we introduce a new realized volatility forecasting technique based on the component structure of the volatility dynamics. Time series of financial volatility is well known for having a complex structure including several heterogeneous patterns: linear and nonlinear, long-run and short-run, etc. We propose a new two-component model of realized volatility that is based on combination of econometric and machine learning approaches. In our model, the ARFIMA framework is used to capture the linear component of realized volatility while the artificial neural network is used to model the corresponding nonlinear part. The model exploits both the strength of ARFIMA in linear modeling and the high nonlinear modeling capability of artificial neural networks. We also develop a modification of the cyclical volatility model where artificial neural networks are used to model both trend and cyclical components of realized volatility. We apply the proposed hybrid approaches to produce out-of-sample forecasts of realized volatility of the GBP/USD and the EUR/GBP exchange rates returns. The proposed models provide an improvement in out-of-sample forecasting accuracy over the competing approaches