Man Institute Catching Angels Before They Fall

Corporate bonds form the cornerstone of most insurers’ strategic asset allocation and typically comprise more than 50% of their balance sheet The ability to anticipate changes in creditworthiness – both downgrades and upgrades – is thus critical to insurance portfolios Research jointly conducted by Man Group, Pension Insurance Corporation (PIC), Stanford University and SAS demonstrates that advanced data analytics may potentially outperform traditional methods to provide a consistent and repeatable process for forecasting rating transitions

For insurers and other regulated financial institutions, balance sheet management is a delicate process. Risk and return must be balanced and financial liabilities matched with appropriate assets that provide both stability and investment income. Corporate bonds play an essential role in this process and typically comprise more than 50% of insurers’ balance sheets. In this context, effective management of credit quality is fundamental to the asset-liability management (ALM) and capital allocation process. Further, the ability to anticipate changes in creditworthiness – both downgrades and upgrades – is critical. A recent research paper, jointly published by Man Group, PIC, Stanford University and SAS, demonstrates that a quantitative credit process, powered by advancements in data analytics and machine learning, may offer more timely insights than traditional approaches. The study provides empirical evidence that machine learning techniques can potentially outperform traditional methods in predicting rating changes, equipping insurers with the insights needed to stay competitive in today's complex credit markets. Here we summarise the key findings from this research.

Before we dive in, it’s worth reminding ourselves of the challenge. Meeting policyholder claims – potentially decades in the future – is the fundamental obligation insurers face. This requires asset portfolios that deliver predictable cash flows and drawdown mitigation. While corporate bonds are ideally suited to meet these needs, their stability is contingent on the issuer's credit quality. For a regulated insurer, an unexpected downgrade is far more than just a mark-to-market loss. Indeed, it has direct consequences across several dimensions:

An unexpected downgrade substantially increases capital charges while simultaneously eroding available regulatory capital – with particularly severe consequences when ratings fall below investment grade. This makes the asset less efficient from a return-on-capital perspective and reduces the insurer's regulatory capital coverage ratio, weakening its competitive position when underwriting new business.

Long-term insurers meticulously match the duration and cash flow profile of their assets and liabilities, especially those operating in the asset-intensive reinsurance market. An unexpected downgrade can result in increased haircuts on asset cash flows under certain regulatory regimes, creating shortfalls in the asset portfolio cash flow required to match regulatory liabilities.

External credit and financial strength ratings assigned to an insurer are fundamental to its ability to transact and underwrite new business. A deteriorating credit portfolio negatively impacts its assigned rating and outlook.

Many insurance credit mandates have strict guidelines, prohibiting holdings of securities below certain credit ratings. A downgrade can thus force portfolio managers to sell assets at inopportune times, often into declining markets, crystallising losses and further impacting the insurer's capital position. Conversely, identifying potential upgrades ahead of the market may allow an insurer to position itself to benefit from credit spread compression and positive price momentum, potentially enhancing portfolio returns without taking on additional credit risk.

In collaboration with PIC, Stanford University and SAS, we explored a machine-learning based framework for identifying firms with an elevated probability of being downgraded or upgraded. Credit markets have historically consumed less data than equity markets, but this is where the leaps we have seen in data availability and investors' ability to process it come into play. Our framework was developed using a proprietary dataset of credit transitions with over 507,000 issuer-month observations between 2001 and 2024. Figure 1 shows the average monthly transition probabilities based on that data. The diagonal blue squares highlight both the infrequency and asymmetric nature of rating changes, demonstrating how challenging it is to forecast these transitions.

Source: ICE BofA IUC0 Index and Man Numeric proprietary data, as of December 2024. Our analysis considered 48 potential drivers (or features) across five different categories that we believe could provide explanatory and predictive power to credit transition: Equity related variables, such as momentum Bond variables such as yield and option-adjusted spread Ratings variables such as current issuer-level rating, split ratedness of the issuer One-year default probabilities from SAS/Kamakura Macroeconomic variables We found strong out-of-sample performance from our machine learning model in detecting downgrades compared to traditional modelling techniques, such as historical transition matrices or linear logistic regression models, when evaluated using ROC AUC (receiver operating characteristic - area under the curve) and precision-recall criteria.1 Note that a ROC AUC of 0.5 represents a random guess while an AUC of 1.0 represents a perfect model. As shown in the top part of Figure 2, our machine learning model achieves an ROC AUC of 0.9, indicating superior performance at predicting credit downgrades. Meanwhile, precision-recall curves show how good a model is at finding rare events (such as credit default of an investment-grade issuer). Precision tells you, “When I flag something as risky, how often am I right?” while recall answers, “How many of the actual risky cases did I catch?”. Precision-recall AUC should be compared to the average occurrence of downgrades, which in our dataset is less than 1%. As shown in the bottom part of Figure 2, the precision-recall AUC for our machine learning model is 12%, representing a 20x improvement over random classification.

Source: Man Numeric proprietary data. Our model identified the following features, shown in Figure 3, as the most influential drivers of credit transitions, based on their contribution to the machine learning model’s predictive performance.

Source: Man Numeric proprietary data. Notes: Importance values are the model-reported relative contributions of each feature. Higher values indicate greater impact on the model's predictions. The importance of option-adjusted spread and yield is well documented, as the bond market tends to price ratings changes well before they occur. Equity market features, such as momentum, also emerge as highly influential, alongside fundamental indicators including default probability and current issuer rating, which serve as critical drivers of transition probability. In contrast, split issuer ratings from rating agencies, a frequently considered feature, were a less significant contributor to future credit migration in our fitted model. The goal in utilising advanced machine learning techniques is not to build a “black box” model that replaces a skilled credit analyst, but rather to create a tool that augments their knowledge. A well-calibrated systematic process may offer more breadth and continuous monitoring of the entire universe of issuers seeking to provide data-driven insight that might help investors identify firms with an elevated possibility of a rating transition in a timely manner. Consistency is another key advantage of this approach. A systematic process evaluates every issuer in the universe against the same rigorous framework, creating a disciplined and repeatable process for credit assessment that allows the insurer to more proactively manage its credit portfolio and capital position.

While machine learning models offer promising advances in predicting credit rating changes, there are also limitations which warrant consideration. Firstly, model performance may not persist across future market regimes or tail scenarios, and false positives risk unnecessary turnover and transaction costs. For insurers with strict mandate guidelines, acting on model-driven signals could also risk crystallising losses in illiquid conditions. To put our model to the test, we worked with our partners to identify two case studies we felt were newsworthy examples of “fallen angels”. It is important to note that, while we believe the results to be accurate, this analysis is provided in hindsight.