When Bed Bath & Beyond filed for Chapter 11 bankruptcy in April, consumers made a mad dash to the business’s storefronts and left with carts of discounted items. The 52-year-old retail company is now in the process of closing 360 Bed Bath & Beyond stores and 120 Buy Buy Baby locations, according to a New York Times article published last month. Bankruptcy news then spilled into early May with Vice Media and Christmas Tree Shops making similar announcements.  

“It’s not often that very large companies go bankrupt so, when it does happen, it’s in the news and it gets a lot of coverage,” says Rachel Cardarelli ’20. But how can companies better predict if their business models will fail?  

Cardarelli, who graduated from Bryant with a degree in Actuarial Mathematics, is a co-author of "Effective Bankruptcy Prediction Models for North American Companies,” a research paper that looks at the imbalance problem in bankruptcy prediction and proposes a new model for North American companies to use. The paper — which she co-authored with Mathematics professors Son Nguyen, Rick Gorvett, and John Quinn — was recently published in The Encyclopedia of Data Science and Machine Learning.  

While bankruptcy prediction research is abundant, few studies deal with the imbalance problem, which the Bryant researchers made their focus. According to the paper, bankruptcy prediction aims to provide risk assessment in order to reduce the likelihood of financial distress for individual companies and the macroeconomy. This information not only gives creditors and investors insight when they must make financial decisions, but it helps them recognize the potential of bankruptcy and how to mitigate costs to stakeholders.   

Nguyen and Cardarelli explain the imbalance problem with the following example: If you have 100 companies, 99 will not go bankrupt and only one will. 

While this information suggests that a model could predict an entirely positive dataset with 99 percent accuracy, Nguyen notes that the negative observations are often ignored since they are so low; however, if the sample size increases, the imbalance worsens.  

“By predicting that everyone's not bankrupt, you achieve a high accuracy and low inaccuracy. It looks like a good model, but it doesn’t actually predict anything because it's assigned everybody not to go bankrupt,” says Nguyen.  

To balance the data, the researchers’ new technique uses an undersampling procedure, which randomly picks a subset of data from non-bankrupt companies and reduces the number of data points so they’re equitable with the number of points of bankrupt companies. Therefore, the proposed technique forces predictive models to focus more on bankrupt companies.  

The Bryant researchers also tested an oversampling technique but found that the method was prone to overfitting, which affects the accuracy of predicting future observations. Nguyen says their proposed technique performs favorably against some of the best and most popular bankruptcy prediction approaches. 

“People can apply this method to their own domain,” he says, adding that businesses can test the proposed method with past data. “The key (to bankruptcy prediction) is not to only go with one method, but to take a look at other methods and try them on the data.”  

The two note that further areas of study could include exploring undersampling techniques and using ensemble techniques to tune balanced ratios in the desired undersampled dataset.  

“We only looked at North American companies, so you could test if this would work in different areas of the world,” Cardarelli says. 

As the shopping carts exiting Bed Bath & Beyond slow in number and Vice Media platforms go dark, more and more companies continue filing for bankruptcy. According to S&P Global Market Intelligence, 236 U.S. corporate bankruptcy filings were made between January and April of 2023 — more than double the comparable figure a year ago and higher than the prior 12 years.