A mixture autoregressive model based on Student’s <i>t</i>–distribution

A new mixture autoregressive model based on Student’s t–distribution is proposed. key feature of our that the conditional t–distributions component models are autoregressions have multivariate as their (low-dimensional) stationary distributions. That with such distributions exist not immediate. Our formulation implies mean each a linear function past observations and variance also time varying. Compared to previous may therefore be useful in applications where data exhibits rather strong heteroskedasticity. has theoretical advantage conditions for stationarity ergodicity always met these properties much more straightforward establish than common nonlinear models. An empirical example employing realized kernel series S&P 500 high-frequency shows proposed performs well volatility forecasting.