Global economic, credit and interest rate cycles are becoming desynchronised — in one market we could see spreads widening and leverage edge up, while in another observe the opposite. Nikko AM’s global credit investment team applies a fundamental-driven investment approach to analyse the global credit market and establish top-down investment themes to benefit from such non-synchronisations. Regarding the former, we rely on fundamental bottom-up analysis when researching individual corporates, analysing balance sheet data, interacting with management teams and predicting cash flows to assess future credit quality of corporates. Nevertheless, beside our strong reliance on fundamental research as an alpha source, we also use quantitative models to aid our credit screening in our investment process. With the global credit universe comprising of over 1,400 corporate bond issuers, full in-depth coverage of each individual corporate represents a challenge to most research teams. To overcome this dilemma, the Nikko AM credit research team has developed a dynamic credit risk model that predicts a corporate’s medium-term default risks (implied by the equity market), fundamental accounting data and historical default events. This enables the investment team to expand to almost full coverage of the aforementioned credit universe. Since its inception, the model has gained wide acceptance among our portfolio managers, partly due to our analysts’ inputs into the development of the model and promising back-test results (see Figure 2).
The credit risk model has been developed for two main purposes.
1) Allows more research resources in the high yield space
Although our model estimates issuer credit ratings implied by the medium term default risk, for corporates along the entire credit quality spectrum, we pay more importance to the investment grade space (BBB-/Baa3 and better) and the crossover area (the intersection of high yield and investment grade bonds). These areas of the credit spectrum are the most informationally efficient, partly due to market’s ability to quickly incorporate new information into prices (and hence default probabilities), and also due to the standardisation of accounting data and extensive coverage by analysts. As such, defaults are relatively easier to model, and therefore forecast, while also being able to achieve the desired balance between type I and II errors. Subsequently, as a significant number of investment grade and crossover issuers get assessed quantitatively, more of the team’s time can be reallocated to the research-intensive high yield space.
2) Acts as a Secondary credit rating provider
The model serves as an additional, independent credit opinion to our in-house fundamental credit assessments. Divergence of results among such often leads to further examination until a final assessment is reached. Just as an academic exercise, we consider firm A – a distressed issuer in the energy sector. The choice of this sector is relevant given the vast majority of defaults globally in 2016 were in that space. Before OPEC’s deal in late 2016 to cut output, the sector was extremely distressed amid a deepening global supply glut which, in turn, sent oil prices plummeting to multi-year lows. Against that backdrop, coupled with deteriorating credit metrics and eroding equity valuations, the model would have exerted downward pressure on firm A’s implied issuer rating, towards CCC/D, well in advance of eventual downgrades by credit rating agencies, which tend to be more reactive than proactive. While the benefits of the model’s early warning signals are clear, analysts’ due diligence remains warranted in order to minimize false signals (type II error).
Since the model has been integrated into our investment process in 2016, it has proven its ability to provide early warning signals for the deterioration in issuers. Details of our model are summarised below.
The Econometric Model
We define a default in our model as, “a corporate’s failure to pay interest/principal on a debt/loan, bankruptcy filing or distressed exchanges.”
Our approach for estimating the probability of default for corporates is to strike the right balance between objectivity, accuracy, stability and timeliness. We achieve this by combining the pioneering work of Altman’s (1968) Z-Score and the economic theory embedded in the Merton (1974) model – a hybrid of structural and statistical models similar to Z-Metrics, Altman et al. (2010). While each of these models has its own strengths and weaknesses, we find that the best approach of default prediction combines equity market and accounting data. That is because neither equity markets nor accounting statements fully reflect all the information related to a company; and the default risk is an empirical exercise better addressed by the data itself as opposed to being specified a priori as in the Merton model. For instance, Falkenstein et al. (2000) find that default probabilities implied by Merton-type models are not consistent with historically observed default rates. Still, Merton’s economic theory offers invaluable insight on how equity value and volatility are interconnected to default risk.
As stated above, fundamental factors and equity market-derived variables are the building blocks of our model. The former are accounting ratios that not only define the creditworthiness of a corporate, but are also distinct from each other and statistically significant in forecasting defaults. These factors are: profitability, earnings volatility, corporate size, interest coverage and leverage. To ensure uniformity across sectors, certain accounting adjustments are applied. In our modelling method the fundamental factors follow those of Altman & Rijken (2004). Specifically, we employ their agency-rating prediction model to objectively determine weights for the fundamental factors. In doing so, we minimise two common forms of estimation problems: omitted explanatory variables and collinearity in the regressor matrix.
Figure 1 (below) compares the distribution of corporates across the rating categories, as rated by S&P and our fundamental model between 1995 and 2015. Overall, the fit is good, although the model-implied rating distribution has a higher kurtosis, reflecting a greater centre of mass around the BB tier (denoted 3 in Figure 1). The differences in shape between the two distributions can in part be explained by variation in the “through-the-cycle” methodology that S&P applies in their rating assessment. More specifically, S&P’s measure of what it treats as a permanent component of default risk varies over time, as it manages the tension between accuracy and stability (its dual objectives).
Figure 1: Distribution of Fundamental & S&P Ratings
Source: Nikko AM
While the fundamental model captures the link between fundamental factors and agency ratings (S&P, see footnote below), the equity model is designed to provide early warning signals for corporates facing near-terms default risks, which builds on the earlier work of Shumway (2001) and Chava & Jarrow (2004). The output from the equity model is statistically derived using a nonlinear technique, which models the nonlinear nature between default risk and three equity market-derived factors. Two of these factors embody the economic theory rooted in the Merton model. According to his theory, the equity holder will drive the corporate into bankruptcy when its asset value falls below its debt level. In this model, the corporate’s equity value and volatility — both of which are observable — are linked to the default risk. The third factor is a risk-adjusted metric of excess return.
Subsequently, we model the outputs of the fundamental and equity models using CreditSights’ methodology by applying multi-year default indicators, which identify corporates that have defaulted over one- to five-year horizons. These estimates are forward default probabilities which are then used to construct a default term structure, similar to Bloomberg’s DRSK and Citigroup’s bank default models.
Let DPt represent the model designed to predict forward corporate default t years from the present, conditional on the given corporate surviving to year t-1. Let i denote the models; then we can represent DPt that we get for years t=1,...,5 as:
Then, for each corporate j, the probability of default in year t conditional survival to year t-1 is given by:
We translate the cumulative 5-year default probability (CDP5,j) into a more familiar metric of credit ratings. The model implied ratings are inferred from a grid that we construct on a periodic basis by establishing the relationship between the median CDP5 within each rating category and the rating itself, as this is less affected by changes in forward default probabilities along the term structure. CDP5,j is computed directly from DPt,j by cumulating survival probabilities over 5 year as:
Finally, corporates are assigned a credit rating as an issuer-level assessment of their credit health using the mapping described above, as summarised in Table 1. All of the values closely match S&P’s global corporate average cumulative default rates (1981 – 2015) for the same time horizon. Robustness of ratings under the model inputs is ensured – to address excessive equity market volatility – by requiring certain conditions to be met prior to a migration. For example, the CDP5 must be moving in the right direction for at least a month (i.e. up for a downgrade and down for an upgrade).
Table 1: CDP for Implied Credit Ratings
Source: Nikko AM
Predictive power test
Figure 2 illustrates the Receiver Operating Characteristic (ROC) curve of our model. The predictive power of a default probability model is characterised by its ability to distinguish healthy corporates from distressed ones. The most commonly used measure of such is the ROC curve and its main statistics – the area under the ROC curve (AUC). We construct the curve by varying the cut-off probability. In particular, for every cut-off, the ROC curve defines the “true positive rate” (percentage of defaults that the model correctly classified as defaults) on the y-axis as a function of the corresponding “false positive rate” (percentage of non-defaults that are mistakenly classified as defaults) on the x-axis. The 45-degree line represents the ROC curve of a random predictor, which has an AUC of 0.5, whereas an AUC of 1 represents a perfect model. Our model has an AUC of ~0.73, which compares fairly well with other available models over the same time period.
Figure 2: ROC curve (1995-2015)
Source: Nikko AM
As it currently stands, Nikko AM’s credit risk model is limited to publicly-listed, non-financial and agency-rated corporates. Although, we could model private entities without equity market-derived components (based purely on fundamental variables), we run the risk of not having an early warning signal of distress. Nevertheless, this issue could be addressed by using sector multiples of public corporates and we might further down the line pursue such an approach to enhance the model. On the financial front, we are currently in the process of building the database to include financial institutions (mainly banks) and defining what constitutes a default in the sector, which is vital in regions where bankruptcies are absent, owing to government bailouts, like in Europe. These model enhancements would aid the team to expand coverage of the universe further.
In this paper, we introduce Nikko AM’s first generation default probability model for corporates. The model incorporates corporate-specific information ranging from equity performance to accounting-based ratios, as well as certain industry effects. The model produces forward-looking default probabilities that provide timely signals of impeding defaults and ratings migration over the credit cycle. Despite the team’s heavy reliance on fundamental research for generating alpha, we believe that quantitative models are vitally complementary to our investment process and focus our team’s time and resources in an efficient manner.
- Whenever possible, we use the S&P actual rating. When the S&P rating was not available, and Moody’s was, we used Moody’s.
- Altman, E.I. (1968). Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy. Journal of Finance, 35, pp. 1001-1016.
- Altman, E.I. & Rijken, H.A. (2004). How Rating Agencies Achieve Rating Stability. Journal of Banking & Finance, 28, pp. 2679-2714.
- Altman, E.I., Rijken, H.A, Watt, M., Balan, D., Forero, J. & Mina, J. (2010). The Z-MetricsTM Methodology for Estimating Company Credit Ratings and Default Risk Probabilities. RiskMetrics Group.
- Bloomberg (2017). DRSK Public Companies. Bloomberg White Paper, Bloomberg LP.
- Chava, S. & Jarrow, R. (2004). Bankruptcy Prediction with Industry Effects. Review of Finance, 8(4), pp. 537-569.
- Falkenstein, E.G., Carty, L.V. & Boral, A. (2000). RiskCalcTM for Private Companies: Moody’s Default Model. Moody’s Investors Service Special Comment.
- Merton, R.C. (1974). On the Pricing of Corporate Debt: The Risk Structure of Interest Rates. Journal of Finance, 29, pp. 449-470.
- Shumway, T. (2001). Forecasting Bankruptcy More Accurately: A Simple Hazard Model. Journal of Business, 74(1), pp. 101-124.
- Sullivan, J. (2010). Bondscore: Managing Default and Ratings Transition Risk across Credit Portfolios. In: IACPM. [online] New York: IACPM, pp. 1-19. Available at: http://web.iacpm.org/dotAsset/16886.pdf [Accessed 30 Dec. 2016].
- Tong, X. (2015). Modeling Banks’ Probability of Default. Applied Economics and Finance, 2(2), pp. 29-51.
- Vasicek, O. (1984). Credit Valuation. KMV White Paper, KMV LLC.