Neural network time series models for financial risk management

This PhD thesis is centered on the use of data mining techniques for improved scenario simulations in a banking asset and liability management environment. Nowadays, there is a proliferation of tools for the automatic construction of predictive or simulation models which may be used for these purposes. A practitioner may be tempted to use them as black boxes to easily build sophisticated multiple time series models, which may later reveal their weakness, showing poor generalization performances. Some important questions are raised here: 1) Should we use „raw“ data or rather resort to an appropriate transformation/filter before modelling? In other words, how should we preprocess the data set? 2) Is there a statistically meaningful way for model selection and assessment? The answers given here, mainly based on linear time series modelling theory and concepts, are neither definitive nor exhaustive, but they proved to be very useful, finally leading to a simple and powerful simulation model.