Analysing Economic Data with Self-Organizing Maps - A Geometric Neural Network Approach
Self-Organizing Maps (SOM) are a special form of Neural Networks that use unsupervised learning and auto-classification of data. Therefore, SOM is a very flexible algorithm which is in particular well-suited to identify unexpected structures in complex (and multidimensional) data sets. We use SOM in order to build feature-domain models, i.e. we rather focus on the geometric or symbolic characteristics of patterns within a time series than on their respective location in time. In a next step we try to extract valuable information from the discovered features in order to forecast out-of-sample. We employ the proposed method with different financial time series and test for its performance by means of a set of non-parametric tests. Moreover, the SOM is employed as a clustering algorithm. The method is used in order to form homogenous groups out of 55 countries only by looking at a set of macro data. Without giving any learning guidelines and/or model restrictions the SOM turns out to be a powerful tool for the identification of clusters in the data through its self-organising behaviour.