Essays in applied econometrics of high frequency financial data

Abstract

In the first chapter, co-authored with Peter Hansen and Asger Lunde, we suggest a novel approach to modeling and measuring systematic risk in equity markets. We develop a new modeling framework that treats an asset return as a dependent variable in a multiple regression model. The GARCH-type dynamics of conditional variances and correlations between the regression variables naturally imply a temporal variation of regression coefficients (betas). The model incorporates extra information from the realized (co-)variance measures extracted from high frequency data, which helps to better identify the latent covariance process and capture its changes more promptly. The suggested structure is consistent with the broad class of linear factor models in the asset pricing literature. We apply our framework to the famous three-factor Fama-French model at the daily frequency. Throughout the empirical analysis, we consider more than 800 individual stocks as well as style and sectoral exchange traded funds from the U.S. equity market. We document an appreciable cross-sectional and temporal variation of the model-implied risk loadings with the especially strong (though short-lived) distortion around the Financial Crisis episode. In addition, we find a significant heterogeneity in a relative explanatory power of the Fama-French factors across the different sectors of economy and detect a fluctuation of the risk premia estimates over time. The empirical evidence emphasizes the importance of taking into account dynamic aspects of the underlying covariance structure in asset pricing models. In the second chapter, written with Bo Laursen, we extend the popular dynamic Nelson-Siegel framework by introducing time-varying volatilities in the factor dynamics and incorporating the realized measures to improve the identification of the latent volatility state. The new model is able to effectively describe the conditional distribution dynamics of a term structure variable and can still be readily estimated with the Kalman filter. We apply our framework to model the crude oil futures prices. Using more than 150,000,000 transactions for the large panel of contracts we carefully construct the realized volatility measures corresponding to the latent Nelson-Siegel factors, estimate the model at daily frequency and evaluate it by forecasting the conditional density of futures prices. We document that the time-varying volatility specification suggested in our model strongly outperforms the constant volatility benchmark. In addition, the use of realized measures provides moderate, but systematic gains in density forecasting. In the third chapter, I investigate the rate at which information about the daily asset volatility level arrives with the transaction data in the course of the trading day. The contribution of this analysis is three-fold. First, I gauge how fast (after the market opening) the reasonable projection of the new daily volatility level can be constructed. Second, the framework provides a natural experimental field for the comparison of the small sample properties of different types of estimators as well as their (very) short-run forecasting capability. Finally, I outline an adaptive modeling framework for volatility dynamics that attaches time-varying weights to the different predictive signals in response to the changing stochastic environment. In the empirical analysis, I consider a sample of assets from the Dow Jones index. I find that the average precision of the ex-post daily volatility projections made after only 15 minutes of trading (at 9:45a.m. EST) amounts to 65% (in terms of predictive R2) and reaches up to 90% before noon. Moreover, in conjunction with the prior forecast, the first 15 minutes of trading are able to predict about 80% of the ex-post daily volatility. I document that the predictive content of the realized measures that use data at the transaction frequency is strongly superior as compared to the estimators that use sparsely sampled data, but the difference is getting negligible closer to the end of the trading day, as more observations are used to construct a projection. In the final chapter, joint with Peter Hansen, Guillaume Horel and Asger Lunde, we introduce a multivariate estimator of financial volatility that is based on the theory of Markov chains. The Markov chain framework takes advantage of the discreteness of high-frequency returns and suggests a natural decomposition of the observed price process into a martingale and a stationary components. The new estimator is robust to microstructural noise effects and is positive semidefinite by construction. We outline an approach to the estimation of high dimensional covariance matrices. This approach overcomes the curse of dimensionality caused by the tremendous number of observed price transitions (normally, exceeding 10,000 per trading day) that complicates a reliable estimation of the transition probability matrix for the multivariate Markov chain process. We study the finite sample properties of the estimator in a simulation study and apply it to high-frequency commodity prices. We find that the new estimator demonstrates a decent finite sample precision. The empirical estimates are largely in agreement with the benchmarks, but the Markov chain estimator is found to be particularly well with regards to estimating correlations.