A1260
Title: Accelerating estimation for GARCH-type models by weighted data subsampling
Authors: Zixuan Wang - University of Technology Sydney (Australia)
Aishwarya Bhaskaran - Macquarie University (Australia)
Thomas Goodwin - University of Technology Sydney (Australia)
Matias Quiroz - University of Technology Sydney (Australia) [presenting]
Abstract: Generalised autoregressive conditional heteroskedasticity (GARCH) models are ubiquitous in financial econometrics to capture and forecast time-varying volatility in asset returns. Thus, they are essential tools in the econometrician's toolbox for risk management, derivative pricing, and portfolio optimisation, and beyond. Likelihood-based estimation of GARCH models is computationally expensive due to the recursive structure of the conditional variance, limiting their applicability in large datasets and scenarios requiring quick decision-making. To speed up the estimation, a subsampling-based unbiased log-likelihood estimator is proposed that includes early observations with a higher probability, thereby reducing the length of the recursive conditional variance loop when evaluating the likelihood. A subsampling-based unbiased gradient of the log-likelihood estimator is also proposed. The proposed estimators may be used in any subsampling-based inferential approach, such as stochastic optimization to find the maximum likelihood estimate in the classical paradigm, or subsampling Markov chain Monte Carlo (MCMC) algorithms to sample the posterior distribution in the Bayesian paradigm.
