Adaptive systems for processing of high-dimensional and multi-scale data: Application to forecasting of geomagnetic substorms and financial time series.
Abstract
Predictive capabilities of the data-driven models of the systems with complex multi-scale dynamics depend on the quality and amount of the available data and on the algorithms used to extract generalized mappings. Availability of the real-time high-resolution data constantly increases in many fields of practical interest. However, the majority of advanced nonlinear algorithms, including neural networks (NN), can encounter a set of problems called "dimensionality curse" when applied to high-dimensional data. Nonstationarity of the system can also impose significant limitations on the size of training set which leads to poor generalization ability of the model. A very promising algorithm that combines the power of the best nonlinear techniques and tolerance to high-dimensional and incomplete data is support vector machine (SVM). We have summarized and demonstrated advantages of the SVM by applying it to two important and challenging problems: substorm forecasting from solar wind data and volatility forecasting from multi-scale stock and exchange market data. We have shown that performance of the SVM model for substorm prediction can be comparable to or be superior to that of the best existing models including NNs. The advantages of the SVM-based techniques are expected to be much more pronounced in future space-weather forecasting models, which will incorporate many types of high-dimensional, multi-scale input data once real-time availability of this information becomes technologically feasible. We have also demonstrated encouraging performance of the SVM in application to volatility prediction using S&P 500 stock index and USD-DM exchange rate data. Future applications of the SVM in the emerging field of high-frequency finance and its relation to existing models are also discussed.
- Publication:
-
APS Division of Plasma Physics Meeting Abstracts
- Pub Date:
- October 2001
- Bibcode:
- 2001APS..DPPKP1078G