Essays on big data, financial risk and forecasting

This dissertation is composed of two separate yet closely linked essays in the field of financial econometrics, dealing with one unifying theme; modeling and forecasting financial risk using high-frequency financial time series data. The first essay aims to shed light on the efficacy of using alternative estimation and testing methodologies related to volatility, jumps, and co-jumps in the financial markets. The second essay on the other hand provides insight into the marginal predictive and economic gains associated with using density and point forecasts for directional volatility prediction. In recent years, the field of financial econometrics has seen tremendous gains in the amount of data available for use in modeling and prediction. Much of this data is very high frequency, and even “tick-based”, and hence falls into the category of what might be termed as big data. The availability of such data has spurred numerous theoretical advances, especially in the areas of volatility (risk) estimation and prediction, and price jump and co-jump detection. Successful risk management, asset pricing, and asset allocation, all depend significantly on asset-return volatility dynamics and discontinuous asset price movements (jumps). Moreover, co-jumps reflect market correlation and are instrumental in identifying systemic risk across multiple sectors and markets. Thus, in the first essay of this dissertation, we begin with a survey of numerous non-parametric estimators of integrated volatility and testing methodologies used to analyze asset-price jumps and co-jumps. Thereafter, to illustrate the finite sample properties of the competing methods discussed in this essay, we carry out an empirical analysis using stock prices of DOW 30 companies and ETFs. With a sampling period of more than a decade, covering the 2008-2009 financial crisis, our empirical investigation reveals commonalities and statistical differences in volatility movement, jump, and co-jump activities as depicted by the different estimators and tests. For instance, jump-robust volatility estimators and the volatility estimator robust to neither jumps nor noise, report similar continuous asset price movements, only when the jump-robust volatility estimator utilizes a data-driven truncation approach to eliminate jumps above a certain threshold. Certain jump detection methodologies are found to indicate more frequent jumps after the financial crisis period (post-2009), thus signifying the over-detection of “small” and “moderate” jumps during the said period. Finally, co-jump tests which utilize results at the intersection of two jump tests, report a higher percentage of days identified as having co-jumps, possibly due to a larger false rejection rate. In the second essay, we provide new empirical evidence of the relative usefulness of interval (density) and point forecasts of asset-return volatility, in the context of financial risk management. In our evaluation, we utilize both statistical criteria, i.e. accuracy of directional volatility predictions, and economic criteria i.e. profitability of trading strategies based on said predictions. We construct interval forecasts using non-parametric kernel estimators, while point forecasts are based on “linear” heterogeneous autoregressive (HAR) models as well as “nonlinear” deep-learning recurrent neural network (RNN) models. Additionally, we utilize a novel trading strategy that builds on the contemporaneous return-volatility relationship and leads to new insights related to linkages between economic and statistical methods of forecast evaluation. Our empirical findings based on high-frequency data suggest that interval forecasts can improve upon point forecasts in terms of trading profitability, regardless of the “linear” or “nonlinear” nature of the point-forecasting model. Moreover, interval forecasts perform better than linear model-based point forecasts, when it comes to directional predictive accuracy. These findings are consistent with hypotheses concerning both nonlinear volatility dynamics and the ability of interval forecasts to accurately estimate “large price jump” induced future volatility movements. We conduct a follow-up series of Monte Carlo experiments, which are motivated by our finding that for translation of statistical improvements into economic gains, the choice of volatility estimation technique is crucial. Our experiments reveal that the inability of certain volatility estimators to accurately predict “pseudo true” volatility density for specific magnitudes of “price jumps” or “microstructure noise” in the price process, can explain why these same estimators are less profitable when used in our empirical trading strategies.Ph.D.Includes bibliographical reference