Bankruptcy Prediction using Robust Machine Learning Model

Abstract

The prediction of bankruptcy is the job of forecasting bankruptcy and different financial crisis measures for businesses. It is an enormous area in business and accounting. The significance of the field is partially attributed to its value for creditors and investors in determining the likelihood of business bankruptcy. A predictive model that combines different economic parameters that enable the financial status of a business to be foreseen is the purpose of predicting financial distress. There were various approaches in this area focused on predictive tests, statistical modeling (e.g. generalized linear models), and in addition, artificial intelligence (e.g. Neural Networks, SVM, Decision Trees). In this work, we record our remarks by designing, experimenting, and evaluating some of the classification models used in most cases i.e. Gradient Boosting, Decision Trees, Balanced Bagging, Random Forests, SVM, and Ada Boost which are applicable to expected bankruptcy. The bankruptcy data is collected from Polish firms, in which synthetic features are used to represent statistics of a higher order. The dataset has outliers and is imbalanced. Synthetic Minority Over-sampling Technique (SMOTE) is used to over-sample minority class labels and tackle the data imbalance issue. The feature selection technique is an important step in the preprocessing in which three techniques were applied i.e. PCA, Select Percentile, and Sequential Feature Selection. To evaluate the models, the results are compared using four matrices i.e. accuracy, F1-score, recall, root-mean-square error (RMSE). The simulation studies reveal that the Ada Boost classifier with SFS as a feature selection method is giving the better result of 98.7% in terms of accuracy.