Niall Murray

Sovereign Credit Ratings Niall Murray November 2018 1 Introduction Following a review of various papers centered around the determination of credit ratings I decided to review Short and Long-run Determinants of Sovereign Debt Credit Ratings written by Afonso, Gomes and Rother. My motivation for covering this paper stems from my interest in the financial markets and the important role which sovereign credit ratings serve. Inflated credit ratings played big role in misleading investors in the financial crisis of 2008, the ten year anniversary of this disaster reminds us of the importance of understanding what actually drives these ratings. Using various types of statistical models from pooled OLS to random effects ordered probit estimation, the authors have uncovered which model is the most suitable for this data and what metrics agencies like Moodys and Fitch tend to use in their models. As well this.... Sovereign credit ratings crucial role in the financial system can broken down as follows: • The primary objective of sovereign ratings is their role in influencing the interest rates at which countries can obtain credit on the international financial markets. This rate is then the basis for which individuals in that country can then borrow and lend at. • Sovereign ratings also influence credit ratings of national banks and companies based in that country. Agencies very rarely assign a credit rating to a bank, company or local municipality that is higher than that of the home country of the issuer (Cantor and Packer, 1996)and affect their attractiveness to foreign investors by directly impacting the ability of firms in that country to access global capital markets. • Many institutional investors can only invest in debt rated at a certain level, any rating below these thresholds may severely limit the flow of capital to a nation. Thus, sovereign credit ratings are strong predictors of a country’s equity market returns and valuations this was documented by (Correa et al. 2014). 1 2 Summary The researchers sought out to analyze what indicators influence sovereign debt ratings and their effects on a short and longer term time horizon. To do this they utilised linear and ordered response models. The linear panel model used is: Rit = BXit + λZi + αi + uit (1) Where Rit represents a linear transformation of credit ratings with AAA=17, Xit is a vector containing time varying variables and Zi is a vector of time invariant variables which include regional dummies. Finally αi stands for the individual effects for each country i. A variety of models such as random effects, fixed effects and pooled OLS were considered to estimate this equation. They had suspicions that the country specific error is likely correlated with the regressors due to factors such as political turmoil, poor infastructure or nations tax policy. There suspicions were confirmed by using a Hausman test where the null hypothesis of no correlation is rejected with p-value of 0.00. With the knowledge that E(αi |Xit , Zi ) 6= 0, one would think you should use fixed effects as both random effects and pooled OLS produce biased and inconsistent estimates under these conditions. Conditional Mean zero assumption has been violated However due to nature of the data, specifically the lack of variation of a country’s rating over time meant that the geographic dummies included in the regression would capture the country’s average rating, while all the other variables would only capture movements in the ratings across time. Essentially leaving the fixed effects regression stripped of meaning. This led the researchers to use random effects and try to deal with the correlation between the αi and the regressors. They highlighted two methodologies to go about this, Hausman-Taylor IV and modelling αi . They decided to model the error term αi as Hausman-Taylor IV estimation would have been a ”Herculean task”.GIVE COMMENTARY They implement this by setting the expected value of the country specific error as a linear combination of time-averages of the regressors X̄i . E(αi |Xit , Zi ) = η X̄i (2) They modify there initial equation (1) with: αi = η X̄i + εi And after rearranging some terms this produces: Rit = B(Xit − X̄i ) + δ X̄i + λZi + εi + uit (3) Where δ = η + B and εi is an error term which is by definition uncorrelated with the regressors. This expression is formulated well as we can think of B and δ as our short term and long term coefficients, B represents the impact of the explanatory variables changing throughout time while δ tracks the long-term effects of our variables. 2 The ordered probit model: ∗ Rit = B(Xit − X̄i ) + δ X̄i + λZi + εi + uit (4) ∗ The unobserved latent variable Rit represents each agencies continuous evaluation of a nations’s credit-worthiness. An ordered probit is a more natural fit for this type of problem, because the rating is a discrete variable and reflects an order in terms of probability of default. The variables c1 to c16 found in (5) represent the threshold parameters, these parameters are limited by the number of credit ratings. This transformation differs significantly from the linear transformation used in the previous models. Utilizing this methodology we get to uncover the shape of ratings curve which is rather illuminating as it tells us whether the system is non-linear or whether one agency may require bigger changes then others. The parameters of equation (4) and (5), specifically B, η , λ and the threshold parameters c1 to c16 are estimated using maximum likelihood. A random effects ordered probit model is used, as it considers both errors εi + uit to be normally distributed. GIVE COMMENTARY This allows one to consistently estimate the parameters. Using previous literature on this subject such as Cantor Packer (1996) and Bissoondoyal-Bheenick (2005), they identified some key variables which may influence sovereign credit ratings. The regressors were then split into four sections. Macroeconomic GDP per capita Real GDP growth Unemployment Inflation Government Government debt Fiscal balance Government effectiveness External External debt Foreign reserves Current account balance Other Default history European Union Regional dummies The ratings data was gathered from Standard Poor’s, Moody’s and Fitch Ratings. In an effort to get more comparable figures fiscal balance, current account and government debt are calculated as a percentage of GDP, 3 foreign reserves are as percentage of imports and external debt as percentage of exports. 3-year averages were taken for the following variables; inflation, unemployment, GDP growth, fiscal balance and current account. The long term average of these figures are used to mirror the rating agencies focus to lessen the impact of the business cycle on a sovereign rating.Unfortunately due to patchy data they end up with an unbalanced panel with 66 countries for Moody’s, 65 for SP and 58 for Fitch, with an average 8 yearly observations per country. Table 1: Results: Significant at 5% level for at least two agencies Short Term GDP per capita Real GDP growth Government debt Fiscal Balance Long Term GDP per capita Default history Government effectiveness Inflation External debt Creating the model with both short and long term coefficients proved to be a success as the vast majority of variables were found to have different effects compared to their long term average. The long term average of variables such as GDP per capita and government effectiveness were found to be significant for each agency which means if they hadn’t included them in the regression they would have suffered from issues of endogeneity. As well as this they used a Hausman test to confirm that the country specific error was now uncorrelated with the explanatory variables. The standout variable form these results is GDP per capita which is significant at 5% level for all agencies across short and long term horizons, indicating what a big role it plays in the determination of these ratings. The random effects ordered probit results were quite similar to that of the linear random effects model with the main difference being additional variables were found to be significant for Fitch, this is likely due to the standard errors being smaller with this model. Ordered probit is more efficient?? Could Fitch having the least data play a role or the fact that this model produces lower standard errors? The key drivers on the short term side were GDP per capita, real GDP growth, government debt which were found to be significant for all agencies. Surprisingly short-term unemployment has a positive impact SP scores at 1% significance level, while long term unemployment has a negative impact for the two other agencies at a significant level which is more in line with my expectations. The short term result seems rather unintuitive to me as unemployment is usually associated with economic downturns, however after further reading a short term rise in unemployment may be a consequence of additional people 4 looking for a job indicating people are positive about the future or that an economy may be experiencing substantial technological innovation which should be followed by prosperous economic growth. The contrast in outcomes between the short and long term coefficients is fascinating and causes one to think more deeply about the specific variable. The assumption that the ratings schedule was linear proved to be reasonable as for the most part the ratings are evenly spaced between each other. SP were found to have highest value for their final threshold parameter indicating an AAA rating from them may be slightly more difficult to obtain. The gap between parameters of junk (BB+) and investment grade bonds(BBB-) is marginally bigger than the average leap between ratings, this is a result I expected and could be explained by the stark consequences of moving between those ratings. Many pensions and other financial institutions are not permitted to invest into debt which is rated below investment grade. The final section and perhaps the most intriguing involves predicting the ratings and the changes of a rating for each agency. The table above highlights the importance of controlling for country-specific effects as we see random effects including the estimated country effect is the most successful model. It correctly forecasted 70 per cent of all observations and more than 95 per cent of the predicted ratings lie within one notch (one notch is difference between two adjacent ratings say AAA to AA+). This result comes as no surprise as I mentioned above factors such as political risk, geopolitical uncertainty and tax policies are contained in this fixed effect term which boosts the accuracy of the model. For me this result points towards the rating agencies focus on including the qualitative variables likely to be contained in the error term when building their proprietary models. 5 Excluding the error term, simple ordered probit came out on top as far as prediction is concerned as it correctly forecasts around 45 per cent of all observations and more then 80 per cent within one notch, random effects ordered probit was slightly behind correctly forecasting 40% and with 78% falling within one notch. COMMENTARY When it came to forecasting a change in a nations rating the models correctly predict between one third and one half of both upgrades and downgrades. Ordered probit and random effects ordered probit predict significantly more changes than the OLS and random effects counterparts. 3 What I liked I liked the use of three agencies, as it allowed one to compare and contrast the methodologies of each firm. Although the results seemed pretty consistent across the board they did diverge on a few variables. The EU dummy is one of them, with Moodys placing a much larger emphasis on it compared to the rest. The formation of the SP threshold parameters was very revealing, the space between each parameter was wider at very top compared to elsewhere on the schedule indicating they require big moves in the fundamentals of a nation to obtain the best ranking. Also we see Fitch place a higher emphasis on average current account deficits and average government balances, its coefficients for these variables are twice the size of their competitors indicating Fitch is more concerned with fiscal responsibility. The process of altering the random effects model was another interesting factor of this paper. I had become accustomed to thinking fixed effects or first difference models were always the way one should deal with correlation between αi and explanatory variables, and they would have went down that road if the data had permitted them but due to stable nature of these ratings they had to think out side the box. RE usually needed the assumption cov(Xit , αi ) = 0, estimating RE without this key assumption was a surprise, and open my mind to how flexible some of these models can be. Another benefit that came with the alterations they made was the level of granularity now offered by model. Splitting the model into short and long term coefficients allows one to see what variables matter and also how they will influence a credit score over time. The current account variable is a good example of this, its short term coefficient is negative while its long term is a large and positive coefficient. So countries who run deficits in the short term are rewarded while those who have a sustained deficits are punished severely, this nuanced information would be crucial for any practitioner involved with investing fixed income. Another feature of this model is that it allows us to somewhat explain the stability of the ratings, we see the long term coefficient of a variable often have greater magnitude then their short term counterpart. If the opposite was the case I think we would see far more changes from the rating agencies. This then brings up the question why do the focus on stable ratings? 6 4 What could be Improved Since there is a large set explanatory which may be correlated with each other the results of these regressions may be adversely affected by multicollinearity. To remedy this issue other researchers such as Teker, Pala and Kent (2013) and Sehgala,Mathurb, Arorab and Guptab (2018) utilised Principal component analysis or factor analysis. PCA allows one to discover if the regressors can be mostly explained in terms of a much smaller number of variables which are called factors. This allows one to reduce the dimensionality of the data set and to identify meaningful underlying factors. In this case it’s possible the researchers were less concerned with the coefficients of the regressions and placed greater emphasis on the prediction results thus they neglected any potential issues which may arise from multicollinearity. I think a greater sample period would allow for increased confidence in their results. During their sample Moody’s was twice as likely to upgrade a rating rather than downgrade, this begs the question is Moody’s overly kind with their ratings or was it the case that this period was time of great economic prosperity and thus downgrades were few and far between. The use of an artificial neural network (ANN) may be better suited to dealing with non-linear relationships, Benell, Crabbe, Thomas and Gwilym (2006) showed that an ANN model is superior to ordered probit model when it comes to modelling sovereign credit ratings as it had increased accuracy when it comes to forecasting. Results like that point towarsd the fact the rating agencies are using neural networks themselves to generate these scores. Use of another model? or any other model provide a better fit? Split the ratings up and see if it performed better for investment grade ratings... 7