Abstract
Using a simple autoregression with exogenous variables (and its transformed error-correction model), we investigate relationships between realized return and risk measured by realized volatility. The empirical results obtained from analysing the German Stock Index (DAX) and the Dow Jones Index (DJ) show a negative relationship between the realized return and risk and also between changes of the realized return and risk for both monthly and quarterly frequencies. There is also some weak evidence of a negative impact of large volatility on changes of the return that can be detected.
The views expressed in this article are those of the author and not necessarily those of the Deutsche Bundesbank.
Notes
1For various definitions of volatilities, see Bollerslev and Zhou (Citation2006).
2See Li et al. (Citation2005) for a literature survey.
3So far, our model is more similar to the Threshold GARCH (TGARCH) model considered by Glosten et al. (Citation1993).
4This reparametrization is commonly used in the context of co-integration and error correction (Banerjee et al., Citation1990). We use this reparametrization for stationary variables, which has a long tradition in econometric modelling in the field of control theory (Davidson et al., Citation1978).
5The TGARCH models capture volatility processes having two-regime threshold dynamics depending on the sign of the shocks. Our empirical results show that such a sign-asymmetry is not significant for any data or any frequency and is therefore excluded in the following theoretical and empirical discussions.
6This phenomenon was already observed in the early 1960s by Mandelbrot (Citation1963) and was denoted as the Joseph Effect that symbolizes a switched persistence between large and small volatilities.
7This data series is the longest DAX series, which covers whole years from its beginning. For a better comparison, we take the DJ for the same period.
8Andrews (Citation1993) proposes a sample range of 70% as optimal to control the trade-off between zero power when and
and zero detection of a possible threshold effect when
.