Local-Linear Estimation of Time-Varying-Parameter GARCH Models and Associated Risk Measures

Abstract

In this article, we propose a nonparametric approach to estimating generalized autoregressive conditional heteroskedasticity (1,1) models with time-varying parameters. We model the time-varying parameters as a smooth function of time and estimate them using a local linear estimator. We show that our estimator is consistent and is asymptotically normal and that the proposed estimator outperforms a rolling window estimator in Monte Carlo simulation experiments. We present strong evidence of parameter instabilities using daily returns of stock indices and explore implications to risk management measures, such as value-at-risk and expected shortfall, through backtesting.