Applying data mining into recognizing signs of financial distress in listed manufacturing and wholesale companies in Vienam turbulent year

Abstract

Our study is to apply data mining in recognizing signs of financial distress in listed manufacturing and wholesale companies in Vietnam’s turbulent year. Due to the economy's volatility, research gaps from domestic papers and Vietnam’s SDG 8 in 2030 about decent work and economic growth, the goal of the study was to (1) expand the applicability of data mining to recognize financial distress in Vietnamese manufacturing and wholesale firms, (2) provide an urgent and early warning signal for these companies, (3) prove that data mining can give better predictive value than multiple discriminant analysis given the same variables, (4) then provide some practical recommendations for managers, investors, creditors, and policymakers. By collecting audited financial statements in 2021, our subject contains 624 listed manufacturing and wholesale companies on three listed stock exchanges: HOSE, HNX, and UPCOM, including a before-preprocessing training dataset (562 companies) and forecast dataset (62 companies). Our new predictive model was built based on Resource-Based Theory. Then, we remove 31 extreme outliers to ensure the good quality of features in the training dataset. By trial and error on 32 different models, we found the most appropriate model and classification methods based on the highest multi-class assessment. To increase the robustness of our new model and classifiers, we (1) built reality in 2022 based on solid criteria excluding financial ratios and (2) compared this with all predictive results from all methods run on the forecast dataset. We achieve our third objective by (1) adopting data mining into the 5-variable model (Altman, 1968) for the forecast dataset, (2) applying again that model using empirical methods, and (3) comparing those two predictive results with the reality benchmark. Our study found that the 7-variable model using the Neural Network method which contains NWCTA, RETA, EBITTA, BVETD, NRTA, LEV, and INTWO gives out the best predictive results among 32 different cases. When comparing based on the reality benchmark, the predictive results from this model have only 2 misclassifications. It also has better predictability than the 5-variable model, given the same classification method. From the predictive results, the mean of each variable in each class is consistent with the theories of financial distress and previous studies. Our study also found that applying data mining has better-predicted results to recognize financial distress than multiple discriminant analyses, given the same variables.