Combining Clustering and Classification Techniques for Financial Performance Analysis

摘要

The goal of this paper is to analyze the financial performance of world-wide telecommunications companies by building different performance classification models. For characterizing the companies' financial performance, we use different financial measures calculated from the companies' financial statements. The class variable, which for each entrance in our dataset tells us to which class any case belongs, is constructed by applying a clustering technique (the Self-Organizing Map algorithm). We address the issue of map validation using two validation techniques. Then, we address the problem of adding new data, as they become available, into a previously trained SOM map, by building different classification models: multinomial logistic regression, decision tree induction, and a multilayer perceptron neural network. During the experiment, we found that logistic regression and decision tree induction performed similarly in terms of accuracy rates, while the multilayer perceptron did not perform as well. Finally, we propose that, with the correct choice of techniques, our two-level approach provides additional explanatory power over single stage clustering in financial performance analysis.