Volatility During the Financial Crisis Through the Lens of High Frequency Data: A Realized GARCH Approach

Abstract

We study financial volatility during the global financial crisis and use the largest volatility shocks to identify major events during the crisis. Our analysis makes extensive use of high frequency (HF) financial data to model volatility and, importantly, to determine the timing within the day when the largest volatility shocks occurred. The latter helps us identify the events that may be associated with each of these shocks, and serves to illustrate the benefits of using high-frequency data. Some of the largest volatility shocks coincide, not surprisingly, with the bankruptcy of Lehman Brothers on September 15, 2008 and Congress's failure to pass the Emergency Economic Stabilization Act on September 29, 2008. The day with the largest volatility shock was February 27, 2007, the date when Freddie Mac announced a stricter policy for underwriting subprime loans and a date that was marked by a crash on the Chinese stock market. However, the intraday HF data shows that the main culprit was a computer glitch in the trading system. The days with the largest drops in volatility can in most cases be related to interventions by governments and central banks.

Quasi score-driven models

“…In this formulation, QMLE also becomes a naturally viable alternative to MLE. Indeed, in a volatility model, large shocks can be downweighted by an appropriate choice of the updating function ρ, e.g., a Student's t log-likelihood, without imposing y t to follow the same conditional distribution; e.g., y t can be assumed to be conditionally Gaussian as in Banulescu-Radu et al (2018). This is in contrast to SD models where there is a strong link between the innovation density and the updating function, which makes it unnatural to use QMLE.…”

“…Our family of models, called quasi score-driven (or QSD) models, therefore encompasses not only SD models but also many other existing models. For instance, Banulescu-Radu et al (2018) use a volatility model whose dynamic is derived from a QSD model obtained from a Student's t log-likelihood to bound the effect of large shocks, although the model is estimated by Gaussian ML.…”

Do high-frequency-based measures improve conditional covariance forecasts?

“…We analyze the one-period ahead forecasting abilities of the RGARCH copula and the bivariate RGARCH models relatively to those of well-established competing models from the literature: the GJR-GARCH copula of Patton (2013) and the corrected-DCC model by Aielli (2013). We also include in this horse race the Robust-RGARCH version of the proposed model by drawing on Banulescu-Radu et al (2018).…”