 Purpose |
New executive information systems for forecasting and optimization
are mathematical and computational challenges. Numerically
intensive computing is coming nowadays to the aid in many
fields of financial services. New modeling, forecasting and
decision supporting systems will become more and more invaluable
tools in finding the best solutions to finance problems. The
range and variety of the underlying information processing
technologies goes far beyond traditional statistical techniques,
such as regression and discrimination analysis, and decision
support systems.
The purpose of our project was to investigate financial markets with nonlinear time series analysis tools. The major aspect concerned the development of uni- and multivariate analysis and forecasting methods for daily financial market data based on feedforward neural networks. Beside this we have also developed several software tools for collecting and archiving data from financial markets all over the world. This concerns daily data as well as high frequency tick-by-tick data. Our data base goes back to February 1994 and is now one of the most comprehensive ones in the world. |
|
|
|
Keyresults |
Connectionist networks of the feedforward type are universal
function approximators. The simplest structures for which this
theorem holds are networks with a single hidden layer and
nonlinear (sigmoid) transfer functions in this hidden layer
only. This theoretical result is accompanied by many numerical
investigations which showed experimentally the capabilities of
relatively simple connectionist networks to approximate nonlinear
mappings.
We followed this concept of nonlinear function approximation using feedforward connectionist networks. Combining the techniques of statistical time series analysis with the ideas inherent in the concepts of connectionist networks we built a feedforward connectionist net approach for modeling and forecasting multivariate time series. |
|
|
|
|
Methods |
The connectionist methodology realized by us and implemented in
our software package includes all stages of the traditional time
series analysis like 1) data preprocessing, 2) model identification
and selection, 3) model building and estimation and 4) methods
for diagnosis checks on its adequacy.
Our connectionist methodology is based on a proper selection of indicators obtained from data preprocessing, a parsimonious network topology with an optimal selected starting solution and a monitoring scheme for the learning process giving us characteristic information on the internal structure and parameters of the network. Furthermore, different nonlinear networks with almost the same performance characteristics can be combined through a decorrelation process with respect to their residuals to achieve in general better forecasts compared to the individual network solutions. Those "antithetic" networks yield much more stable predictors and allow to specify error bounds which are an essential step forward towards a confident network design. |
|
|
|
|
Validation |
The connectionist network methodology was implemented in form of
a statistical software package. Besides its C and FORTRAN program
library which contains state-of-the-art methods and algorithms,
the software package embodies a simple command language to design
new applications and a graphical user interface to visualize
complex system behavior.
The connectionist time series approach was applied to a series of financial and economic data including for example exchange currency rates, interest rates, and electric power consumption. |
|
|
|
|
Perspectives |
Beyond the connectionist feedforword approach we have investigated
several other nonlinear time series approaches, especially
designed for intraday financial market analysis, forecasts and
risk estimation.
The project offers also the possibility to commercialize our research activities, i.e. to design and develop a commercial statistical software package for nonlinear time series analysis. |
|
|
|
|
Contact |
PD Dr. Diethelm Würtz Swiss Center for Scientific Computing ETH Zentrum, CLU B3 8092 Zurich, Switzerland e-mail: wuertz@scsc.ethz.ch Tel. +41-1-632.5567 Fax. +41-1-632.1104 |
|
|
|
|
Partners |
The project was realized within a series of cooperations. Especially, we thank Swiss Bank Corporation in Basel and Zurich and Reuters in Zurich for their continuous help and support during this project. |
|
|
|