بررسی پیش بینی ورشکستگی در مجموعه داده های نامتعادل An investigation of bankruptcy prediction in imbalanced datasets

1. Introduction The most recent financial crisis exposed the vulnerability of the financial system; more than ever before, firms of all sizes are suffering financial difficulties that sometimes lead to bankruptcy. Such difficulties affect financial institutions, shareholders, managers, employees, and governments alike, and it is crucial to be able to predict corporate bankruptcy. In turn, this critical corporate issue has become a major research area in the corporate finance field. Although several corporate bankruptcy prediction models have been proposed, according to various prediction methods or variables (Balcaen and Ooghe, 2006), most have been designed using the classical paradigm of paired samples of available data (Chen et al., 2009; Olson et al., 2012). That is, the datasets contain the same number of bankrupt and nonbankrupt firms. Such a practice ignores real-world conditions, where bankruptcy is rare. Although the number of nonbankrupt firms is high, the proportion of bankrupt firms is very low, on an order ranging from 100:1 to 1,000:1. Therefore, in the real world, researchers face imbalanced datasets, in which bankrupt company observations are clearly outnumbered by non-bankrupt companies. Therefore, we explore the predictive capacity of bankruptcy models in imbalanced datasets. Data and their characteristics are the most crucial elements of any prediction model (Anderson, 2007), so the imbalanced class distributions in datasets are relevant and demand analysis. The issue of data imbalance has been documented from two perspectives. The first acknowledges that when a bankruptcy prediction model uses a dataset that represents the real-world population – that is, an extremely low frequency of firm of firm bankruptcies- model’s predictive performance is diminished, especially for bankrupt firms. The second offers a treatment technique for handling imbalanced datasets and improve the model’s classification accuracy. Although these perspectives provide a foundation for understanding this issue, fundamental questions remain. Which imbalanced class distribution disturbs a model’s predictive performance? What is the improvement capacity of treatment techniques? Datasets may present multiple imbalanced class levels that contain different proportions of bankrupt firms, because of the irregular bankruptcy rates in the population, the scarcity of bankrupt firms, and a lack of accessibility to these firms’ information (Tian et al., 2015). To evaluate whether a bankruptcy prediction model’s forecast capacity is jeopardized, it is essential to address the imbalanced proportion that significantly disturbs the performance of the model. Moreover, given that bankruptcy is a critical corporate issue that has social costs, it is important to predict it accurately. We therefore conduct an analysis of the capacity of treatment methods to predict bankruptcy in a scenario marked by imbalanced datasets.