In this paper we analyze an econometric model for non-stationary asset returns. Volatility dynamics are modelled by nonparametric regression; consistency and asymptotic normality of a symmetric and of a one-sided kernel estimator are outlined with remarks on the bandwidth decision. Further attention is paid to asymmetry and heavy tails of the return distribution, involved by the framework for innovations. We survey the practicability and automatization of the implementation. For simulated price processes and a multitude of financial time series we observe a satisfying model approximation and good short-term forecasting abilities of the univariate approach. The non-stationary regression model outperforms parametric risk models and famous ARCH-type implementations.
heteroscedastic asset returns non-stationarity nonparametric regression volatility innovation modelling forecasting Value at Risk (VaR) ARCH-models