Evaluating Machine Learning and Deep Learning Analytics for Predicting Bankruptcy of Companies

Corporate bankruptcy is a global issue that has increased over the last few years. Due to lack of adequate historical data, current models have not been able to correctly predict cases of bankruptcy. This research proposed a composite procedure at four stages which includes pre-processing and data rebalancing methods to curate the data, perform feature selection, use various machine learning and deep learning models to construct a robust predictive bankruptcy model and the use of explainable AI to understand the various features that contribute to the model prediction. Models were based on the labelled historic data of various bankrupted and non-bankrupted Polish companies between the period of 2007–2013. Prior to models’ application, data were split to training and test sets that consisted of 25,122 and 12,375 datapoints, respectively. Models were evaluated on various metrics: ROC-AUC, recall, and F-Beta score to determine the best predictive model. Our comparative study showcases that missing data imputation performed using KNN Imputer, skewness reduction using Yeo-Johnson transformation, feature elimination using Recursive Feature Elimination technique along with cost sensitive learning used in tandem with XGBoost algorithm produces the best model with test AUC score of 96.1%, recall score as 96%, and F-Beta score of 92.42%. Implementation of Explainable AI has also helped in realizing top four significant features that impact negatively to bankruptcy prediction globally and locally across all the models created are the ratios “total_cost_overtotal_sales”, “gross_profit_in_3_years_over_total_assests”, “profit_on_sales_over_sales”, and “profit_on_sales_over_total_assests”. Such insights on the classification outcome which instils confidence amongst the decision makers about the validity of the model and its prediction capabilities.