A test of GARCH models onCoCo bonds

Sammanfattning: This research investigates to what extent the ARCH model and the GARCH model forecasts one-day-ahead out-of-sample daily volatility (conditional variance) in European AT1 CoCo bonds compared to the Random Walk model. The research also investigates how different orders of ARCH and GARCH models affect the forecasting accuracy. Specifically, the models investigated are the Random Walk model, ARCH(1), ARCH(2), ARCH(3), GARCH(1,1), GARCH(1,2), GARCH(2,1), and the GARCH(2,2)model. The data set used in this report is 47 European AT1 CoCo bonds from 20 different issuers.The results show that 42 out of 47 CoCo bonds have daily log returns that are conditional heteroscedastic. Five CoCo bonds with homoscedastic daily log returns were CoCo bonds with significant low liquidity. The results show that the GARCH model outperforms both the Random Walk model and the ARCH model, under the assumption that the innovations follow a normal distribution. The results also show that a higherorder of ARCH or GARCH does not necessarily lead to more accurate forecasts. The GARCH(1,1) model provided the most accurate predictions. The conclusion is that the GARCH models provide accurate volatility forecasts in CoCo bonds compared to the ARCH-model, and the Random Walk model. However, the ARCH model and the GARCH model fail to forecast the daily volatility in CoCo bondswith insufficient liquidity. Furthermore, a higher order of ARCH or GARCH models does not necessarily lead to better forecast results.