Dresden 2003 – scientific programme

Parts | Days | Selection | Search | Downloads | Help

AKSOE: Physik sozio-ökonomischer Systeme

AKSOE 1: Finanzm
ärkte und Risikomanagement I

AKSOE 1.4: Talk

Monday, March 24, 2003, 11:00–11:30, BAR/205

Neural Networks for Volatility Predictions — •Friedrich Wagner — Institut für theoretische Physik, Universität Kiel, Leibnizstr., D 24098 Kiel

Neural nets can be used to predict the volatility averaged over one month in the period 1990-2002 from the data in the years 1974 -1989. Important for the application is the choice of the input data and less important the net work architecture. An approach with neural nets has the nice feature that upper and lower bounds can be constructed. Technically the serious problem of over fitting has to be solved.

The quality of the predictions is slightly better than the those from a GARCH(1,1) model,which is considered as a bench mark model by economists. Also the time structure of the volatility cluster is studied and compared with the result from herding models.