Assessing the Long-Term Credit Standing of Debt Issuers Using Dimensionality Reduction Techniques Based on Neural Networks An Alternative to Overfitting

摘要

The purpose of this paper is twofold: first, to propose a new quantitative system to assess the long-term credit standing of debt issuers using neural networks; and second, to propose a new methodology to avoid the problem of overfitting by applying linear and non-linear dimensionality reduction techniques based on neural networks. After experimentation, we found that neural network implementations of linear Principal Component Analysis (NN-PCA) and non-linear PCA (NN-NLPCA) explain a higher proportion of variation in the original set of variables than the common Principal Component Analysis (PCA) methodology. The Principal Components (PCs) extracted from NN-PCA and NN-NLPCA are better discriminators than the PCs extracted from PCA and are easier to interpret if extracted from homogeneous groups of financial ratios. The results of this study suggest that linear and nonlinear dimensionality reduction techniques based on neural networks can be an efficient tool to assess the long-term credit standing of debt issuers and at the same time provide an alternative solution to overfitting.