Bayesian Semi-Parametric Realized Conditional Autoregressive Expectile Models for Tail Risk Forecasting

A new model framework called Realized Conditional Autoregressive Expectile is proposed, whereby a measurement equation is added to the conventional Conditional Autoregressive Expectile model. A realized measure acts as the dependent variable in the measurement equation, capturing the contemporaneous dependence between it and the latent conditional expectile; it also drives the expectile dynamics. The usual grid search and asymmetric least squares optimization, to estimate the expectile level and parameters, suffers from convergence issues leading to inefficient estimation. This article develops an alternative random walk Metropolis stochastic target search method, incorporating an adaptive Markov Chain Monte Carlo sampler, which leads to improved accuracy in estimation of the expectile level and model parameters. The sampling properties of this method are assessed via a simulation study. In a forecast study applied to several market indices and asset return series, one-day-ahead Value-at-Risk and Expected Shortfall forecasting results favor the proposed model class.