Machine learning model gives early warning on credit rating shifts

A new machine learning-based corporate credit risk model has demonstrated significant improvements in forecasting credit rating changes, according to results from a collaboration between data and AI firm SAS, investment manager Man Group, Pension Insurance Corporation, and Stanford University.

The model aims to provide early warning indicators for credit rating upgrades and downgrades, allowing investors and portfolio managers to act before such changes are fully priced by markets and reflected by ratings agencies.

Model design

The model incorporates over twenty years of historical data, including KRIS default probability metrics, bond spreads, yields, equity performance indicators, and macroeconomic variables. Utilising machine learning techniques, it predicts the likelihood of a company's credit rating being upgraded, downgraded or remaining unchanged. Backtesting indicated that this approach ranked firms more accurately by future rating changes compared to traditional risk assessment tools.

Early-warning potential

The system is positioned as a forward-looking risk management aid, offering signals ahead of market pricing. This is especially relevant for investors and asset managers sensitive to credit ratings, such as insurers and fund managers with regulatory investment-grade mandates. More accurate prediction of rating transitions is expected to assist in capital allocation and portfolio management decisions.

"The breakthrough approach of our new model shows that investors can do a lot better than the current best practices. The model's early-warning signals give them critical time to act before the market fully prices in the event, helping better manage risk, reduce losses and seize opportunities," said Stas Melnikov, Head of Quantitative Research and Risk Data Solutions at SAS and a member of the team that developed the new model.

Stress in credit markets

The timing of the model's introduction comes amid higher borrowing costs and pressure on corporate credit profiles. While the Bank of England has maintained interest rates at 4%, refinancing costs remain elevated compared to recent years, placing strain on companies needing to roll over debt and increasing the risk of defaults. The model is designed to flag potential rating changes before asset prices fully reflect such risks, which can influence the availability and cost of credit for companies and the valuation of existing corporate bonds held by institutional investors.

Data features

The research drew on more than 500,000 records covering credit events from 2001 to 2024. The study found KRIS' one-year default probability (KDP) to be the third-most influential factor in the machine learning model's predictions, after option-adjusted spread (OAS) and yield-to-maturity (YTM). According to the team, the use of KDP contributed information that helped identify companies at risk of rating transitions in advance of pricing by the market.

Bond ratings play a pivotal role in financial markets. If a company's debt is downgraded below investment-grade, certain institutional investors are required by policy to sell such holdings, which can trigger price declines for affected securities. The early identification of such risks can give asset managers more flexibility to adjust portfolios proactively.

Industry applications

Insurance companies and asset managers are among those expected to benefit from the model's insights, as they are subject to regulatory and risk constraints tied to credit quality. As the new tool is integrated with SAS's suite of risk management offerings, it could supplement asset and liability management, credit risk oversight, and provisioning for expected credit losses.

Steven Desmyter, President of Man Group, commented on the model's performance: "The results surprised even us. The new models significantly outperformed traditional approaches. In plain English, they were much better at telling us which companies were at risk of downgrade or upgrade."