Bank of America Corp. has begun using artificial intelligence to predict the likelihood of companies defaulting on loans.
“Today we present our inaugural work on applying the latest machine learning tools to analyzing the credit risk,” Bank of America credit strategists Oleg Melentyev and Eric Yu and head of predictive analytics Toby Wade said in a research note Friday. They have started using natural language processing to digest earnings-calls transcripts in order to estimate companies’ probability of default over the next 12 months.
In expanding their default model with the help of AI, the credit strategists seek to detect language used by chief executive officers and chief financial officers that signals a company’s high likelihood of default. Phrases that link to defaulting include cost cutting, asset sales, and cash burn, they said.
Natural language processing has pointed to “more significant credit stresses” in sectors exposed to Covid-19 than under Bank of America’s existing default framework, according to the note. For example, the machine-learning technology predicts default rates will be higher in energy, transportation, and media, and lower than estimated in the cable and health-care sectors.
“This self-learning nature of the NLP signal could make it indispensable to future modeling efforts in credit,” Melentyev, Yu, and Wade said in their note. “We intend to keep exploring this exciting new area in our research in coming months.”
[II Deep Dive: AI-Powered Hedge Funds Vastly Outperformed, Research Shows]
For now, Bank of America is keeping its existing default-rate model along with its forecast for a 5.75 percent default rate over the next 12 months.
“We are planning to observe the behavior of the new BofA NLP-driven US HY Default Rate Indicator independently and compare the estimates produced by the existing model against one that incorporates it,” the bank’s strategists and head of predictive analytics said in the note.
The use of AI increases Bank of America’s forecast for defaults by 15 basis points to 5.9 percent.
The widest deviation in default rate based on sector is in energy, with natural language processing driving the bank’s estimate to 18 percent from 11 percent. In other industry gaps, NLP signals drove forecasts for defaults in transportation to 11 percent from 10 percent, and in media to 8 percent from 4 percent, according to the note.
“While most phrases identified by the NLP model seem to make intuitive sense, not all of them are doing so,” Melentyev, Yu, and Wade said. Phrases like “investor relations,” “oil gas,” and “cash generation” probably result from “spurious correlations picked up by the engine.”
But they’re not rushing to intervene at this point.
“At this stage of our research we decided to keep all phrases without intervening and flagging some of them as false signals,” they said. “This area represents an area of further research, where the human expert input is likely to make a purely machine-learning algorithm produce better results.”