MEAFA workshop on Neural Networks in Econometric System Identification, Forecasting and Control, 15-17 Feb 2010
- Dr Hans Georg Zimmermann, Senior Principal Research Scientist, Siemens AG - Corporate Technology, Germany
Standard econometric tasks that are used in the identification of dynamical systems, their forecasting and control are an important part of optimal decision making. The questions of whether neural networks add something new to the optimal decision making process, as well as providing new methodological insights and better understanding of the real world economy are answered positively in the course of this workshop.
Developments in recent years have provided a deeper understanding of appropriate neural network architectures and corresponding learning algorithms which are beyond pure data-driven modelling. With these techniques, new types of real world applications such as long term forecasting, or the control of dynamical systems can be attacked.
Given the availability of a sufficiently large amount of data, the system identification problem is mainly a learning task. But how should we proceed if the available data set is relatively small in relation to the underlying problem complexity? One possible answer is the integration of additional prior information which can be used as a pre-design of the network architecture. Beside this, the modelling procedure itself can be extended: Instead of insisting on the model building certainties (data, priors) we can measure the uncertainties and use these information to control the learning. This gives us the chance to work with models that are underestimated from the viewpoint of standard regression theory.
Since today's financial markets are highly interrelated, the analysis of an open dynamical system is not sufficient. What is required is to model coherent market movement as a closed dynamical system. Here, we understand a part of the world as a large recursive system which is only partially observable. We model and forecast all observables, avoiding the problem in open systems where we do not know the external drivers from present time onwards. This framework goes far beyond the paradigms of standard regression theory and allows us to perform a new approach to risk analysis which is especially important in finance.
The workshop is of interest to both academics and practitioners. It gives an overview of 22 years of development of the method and software used by Siemens Corporate Technology, together with financial econometric applications.
All days run from 09:00-17:00 with morning break, lunch and afternoon break. Catering will be provided.
Monday 15 Feb 2010
- Introduction to Neural Networks: A perspective from Biology, Economics and Mathematics
- A Correspondence Principle for Neural Networks: Learning - Gradient based learning and the Observer Observation Dilemma
- A Correspondence Principle for Neural Networks: Learning - Occam and Bayesian methods
- Whitening the black box: Sensitivity Analysis and Feature Selection.
Tuesday 16 Feb 2010
- Recurrent Neural Networks - a correspondence principle: The modeling of open dynamical systems, including Error Correction Neural Networks
- Recurrent Neural Networks: Learning of open dynamical systems
- Variance - Invariance Analysis and dynamical systems on manifolds
- Conditional density analysis with feedforward and recurrent neural networks
Wednesday 17 Feb 2010
- Long term Forecasting and modeling with large recurrent neural networks: Beyond the paradigms of regression theory
- Learning of large recurrent neural networks
- A new concept of forecast uncertainty in forecasting and risk and finance
- Optimal control of observed dynamical systems
The workshops will take place at the Faculty of Economics and Business computer Lab 2, Building H69 ground floor, cnr Codrington & Rose streets, the University of Sydney, see interactive map.
- The workshop is fully subsidised for MEAFA members and graduate and higher research degree students of the University of Sydney
- A fee of $300 (excl. GST) applies to other member of staff from the University of Sydney, in order to cover for the daily running costs
- A fee of $900 (excl. GST) applies to all other academics and practitioners outside the University of Sydney
For more details and to reserve a place at the workshop contact email@example.com