Trend Prediction of Exchange Rate with Regularization Networks on Dimension-Adaptive Sparse Grids
Description
The prediction of trends from historical market data is an important tool for
decision making in short and long-term investions. The classical time series
analysis cannot reproduce the complex, nonlinear behaviour especially of
coupled time series. Nonlinear approximation methods such as neural networks
and data mining algorithms cannot be used for efficiency reasons due to the
extremely large number of data points and the high dimension of the attribute
space.
Therefore, we want to use for trend prediction an approach which is based on
regularization networks and which uses sparse grids. This approach scales
linearly with the number of data points and allows a nonlinear data analysis.
Furthermore, the uniqueness of the attribute function is not required. In
addition, dimension-adaptive refinement allows an automated dimension reduction
of the problem and, this way, a fast computation.
Partners
This project is funded by the University of Bonn as part of the REMO
(complex retrieval and monitoring scenarios in science and engineering)
initiative.
References
-
T. Gerstner, M. Griebel,
Dimension-adaptive
tensor-product quadrature, Computing, 71(1):65-87, 2003.
-
J. Garcke, M. Griebel and M. Thess,
Data mining with sparse
rids, Computing, 67(3):225-253, 2001.
-
The Sparse Grid
Bibliography.
Related Projects
