Comparison of Support Vector Machine and Back Propagation Neural Network in Evaluating the Enterprise Financial Distress

Recently, applying the novel data mining techniques for evaluating enterprise financial distress has received much research alternation. Support Vector Machine (SVM) and back propagation neural (BPN) network has been applied successfully in many areas with excellent generalization results, such as rule extraction, classification and evaluation. In this paper, a model based on SVM with Gaussian RBF kernel is proposed here for enterprise financial distress evaluation. BPN network is considered one of the simplest and are most general methods used for supervised training of multilayered neural network. The comparative results show that through the difference between the performance measures is marginal; SVM gives higher precision and lower error rates.

💡 Research Summary

The paper investigates the application of two prominent machine‑learning techniques—Support Vector Machine (SVM) with a Gaussian radial basis function (RBF) kernel and a Back‑Propagation Neural Network (BPN)—to the problem of enterprise financial distress classification. The authors begin by highlighting the economic significance of early distress detection for managers, investors, and credit institutions, and they review prior work that employed artificial neural networks (ANNs) and SVMs for similar tasks. While ANNs are praised for modeling nonlinear relationships, the authors note their susceptibility to local minima, slow convergence, and over‑fitting, especially when training data are scarce. In contrast, SVMs are presented as robust classifiers that achieve global optimality through structural risk minimization, even with limited samples.

The theoretical section on SVM details the construction of a separating hyper‑plane, the formulation of the primal quadratic programming problem, and its dual representation via Lagrange multipliers. The authors explain the kernel trick, focusing on the Gaussian RBF kernel (K(x,x’) = \exp(-\gamma |x-x’|^2)), and they describe the practical steps for model building: data formatting, scaling, kernel selection, and hyper‑parameter tuning (C and γ) using 3‑fold cross‑validation. Implementation is carried out with the LIBSVM library; the optimal parameters are found to be C = 1 and γ = 0.25, yielding nine support vectors for the training set of 20 instances. The resulting model achieves perfect classification (100 % accuracy), zero mean‑square error, and a correlation coefficient of 1, indicating an exact fit to the training data.

The BPN portion outlines a multilayer feed‑forward network with a 4‑4‑1 topology: four input nodes representing aggregated financial ratios, four hidden neurons, and a single output neuron that encodes distress (+1) or non‑distress (‑1). The learning rate (α) and momentum (η) are tuned empirically within ranges 0.6‑0.9 and 0.1‑0.4, respectively. Training is performed using MATLAB’s neural‑network toolbox, employing the standard back‑propagation algorithm to minimize the root‑mean‑square error (RMSE). After 258 iterations, the network reaches a classification accuracy of 95 % (19 correctly classified out of 20 training cases). The authors also discuss a heuristic for determining hidden‑layer size, (h \approx \frac{m}{5(m+n)}), where m is the number of training patterns and n the number of output nodes.

Data for the empirical study are drawn from a Taiwanese securities firm’s database, comprising 45 listed companies: 20 for training and 25 for testing. Fifteen raw financial indicators (e.g., debt ratio, ROA, inventory turnover) are aggregated into four composite scores—financial structure, earning ability, operating ability, and debt‑paying ability—using expert‑assigned weights and a simple averaging formula. The resulting feature vectors are normalized before being fed into both classifiers.

Experimental results compare the two models on the same dataset. SVM attains 100 % accuracy on the training set and, by implication, on the test set (the paper reports only training performance). BPN, despite a reasonable architecture, falls short with 95 % accuracy and a higher number of training iterations, reflecting slower convergence and sensitivity to hyper‑parameter settings. Error analysis distinguishes Type I (false negative) and Type II (false positive) errors; SVM records none of either, whereas BPN exhibits a few Type I errors.

The authors conclude that SVM, even with a modest sample size, delivers superior precision and lower error rates than a conventional back‑propagation network for financial distress prediction. They acknowledge limitations such as the small sample size, the subjective weighting of financial ratios, and the lack of feature‑selection or dimensionality‑reduction techniques. Future research directions include expanding the dataset, exploring multi‑class scenarios, integrating advanced preprocessing (e.g., principal component analysis), and investigating ensemble methods that could combine the strengths of SVM and neural networks for more robust enterprise risk assessment.