The name is absent

random effects have a propensity to overestimate ratings that are bellow CCC+ (Caa1),
on the other hand, both ordered probit prediction errors are quite balanced.

[Insert Figures 3, 4 and 5 here]

Those figures provide some additional insights. For Moody’s, ordered probit performs
well in the bottom ratings while random effects ordered probit is better for top ratings.
Also notice that all four models have difficulty explaining the rating A3. Out of 21
observations the maximum correctly predicted is 2 (with OLS) with a substantial
number of predictions lying outside 1 notch.

For S&P the ordered probit outperforms all other models in the middle and bottom
categories. For Fitch, one should mention that the number of observations used for the
random effects ordered probit is higher than the other models (because of the non-
inclusion of one of the variables), which makes comparison harder. One element we
need to highlight is the fact that there is only one observation in the category A+, which
is the possible cause for the identification of the jump in both limited dependent
variable estimations, mentioned before in section 4.4. For completeness, Figure 6
reports the map of predicted ratings using the random effects estimation including
country specific errors.

[Insert Figure 6 here]

Let’s now turn to how the models perform in predicting changes in ratings. Table 16
presents the total number of sample upgrades (downgrades), the predicted number of
upgrades (downgrades) and the number of upgrades (downgrades) that where correctly
predicted by the several models.

[Insert Table 16 here]

Roughly the models correctly predict between one third and one half of both upgrades
and downgrades. In our opinion this is quite satisfactory given that the empirical
approach used here necessarily neglects two sources of information that are known to
enter the decision of the rating agencies. First, in contrast to the backward-looking

25

<<   22   23   24   25   26   27   28   >>

More intriguing information