Causality Versus Predictability In Neural Network Modeling

In this talk I want to understand Causality as the intellectual effort to interpret a model behavior (e.g. a forecast) it is the answer to the question `Why something happens . On the other hand we have the goal to compute best possible predictions it is the answer to the question `How something happens . Unfortunately both goals do not match. To see this, we have to do a fast review on different forecasting methods. Neural networks are an appropriate framework for the modeling of high dimensional, nonlinear models. This may be function approximations or state space models, realized in form of various recurrent neural networks. Along this line we will improve our predictability, but loose at least a part of the interpretability. This is not a drawback of the modeling but a result of reconstructed unobserved hidden variables in the more advanced models. It is our decision to focus on the WHY or on the HOW.