Stevens Researchers Build Deep-Learning Model That Reads Between the Lines to Predict Bankruptcy
Professors Say Collaborative Culture at the School of Business Led to Innovative, Interdisciplinary Research
When Paul J. FitzPatrick published his study comparing the accounting ratios of successful and failed companies from 1920 to 1929, he said the amount of literature on the topic was meager.
That was in 1932. Today, bankruptcy prediction is among the most robust areas of research in the business world, with thousands of studies dedicated to forecasting corporate financial distress.
So how were two researchers from Stevens Institute of Technology able to build a more accurate model?
By thinking outside of the typical business-school box.
“Most of the other studies used numeric data for prediction, but our model takes the traditional numerical data and combines it with the textual data each public company files annually,” said Dr. Chihoon Lee, an associate professor at the School of Business who collaborated with assistant professor Dr. Feng Mai for this study, published in the European Journal of Operational Research.
They used deep-learning techniques to analyze text within the Management Discussion & Analysis section of annual filings for nearly 12,000 publicly traded U.S. firms from 1994 to 2014.
Although other models have included text, Dr. Lee said none have been as comprehensive as this model.
“Those models just count the important words,” he said. “They don’t go as deep into the meaning of words. We go beyond that and extract more information, which improves prediction power.”
Extracting information from the MD&A text wasn’t without its challenges. The researchers had to find a way to quantify the unstructured textual data in a way that a machine could understand.
“The beauty of a neural network is that it can automatically learn more abstract representations of the words,” Dr. Mai said. “The meaning of the word is condensed into a list of numbers, which only the machine can understand. But it turns out that list of numbers captures the meaning of that word really well.”
Finding new research areas to explore
This research marks a convergence of parallel paths for professors Lee and Mai, who both began teaching at Stevens in the fall of 2015. Although neither had a background in finance, both found themselves drawn to the field.
“I didn’t have too much experience in finance-related research,” said Dr. Mai. “My main area is computational methods, especially machine learning, to analyze text data. But finance is one of the pillars of the School of Business, and I started collaborating with those professors frequently.”
Dr. Lee, who focuses on using statistics in operations research to optimize systems under uncertainty, said he left Colorado State for Castle Point because “[he] could do more impactful research here.”
“Before Stevens, my work was more about proving mathematical properties on certain stochastic systems,” said Dr. Lee. “Now I’m thinking beyond that — not only mathematical properties, but real implications.”
The financial crisis of the late aughts was the most recent reminder of just how real bankruptcy’s implications are. Their model is a way to help shareholders and investors interpret more of the information provided by companies in order to make smarter decisions.
“I think it can help the efficiency of the market if investors have a more accurate way of understanding the textual documents and realizing what kinds of risks are out there,” Dr. Mai said.
Bold moves and unconventional methods
They hope to continue working together to improve the model. One avenue they are exploring is the possibility of adding another, ostensibly more objective, layer of textual data in the form of news. It’s the sort of research risk-taking the School of Business fosters.
“There’s a collaborative culture at the School of Business that promotes interdisciplinary, tech-driven research,” Dr. Mai said. “Two professors from distinct backgrounds can work together because there isn’t a rigid departmental structure. And the leadership here encourages us to make bold moves while using untested methods on important problems.”
By working together, Stevens faculty are able to glean valuable insights and learn about the latest methods from each other. This project marked Dr. Lee’s first foray into deep learning, while Dr. Mai had never used deep learning to process financial data. The novelty of this prediction model is a byproduct of the novel tools that researchers at the School of Business are eager to explore.
“Deep learning is definitely not in the traditional toolbox of a business researcher,” Dr. Mai said.
Then again, Stevens isn’t a traditional business school.