A post on The Guardian’s Datablog triggered my curiosity. While the post itself is basically an explanation of the work of Moody’s, S&P’s and Fitch’s on rating Souvereign credit, it came with an interesting set of historical data. The spreadsheet contains data from over a year of credit ratings per country.
Naturally, scoring a country’s credit is an art in itself, taking into account multiple economical and political factors. But what does today’s rating say about tomorrow’s? Is there predictive power in this data?
To find out, I took the data that The Guardian supplied. I collected all periods in a single file, showing per country, per date the rating and outlook of each agency. Then I added to every row the country’s rating and outlook of each agency from the previous period. Next I created a model to predict the S&P rating, based on the ratings and outlooks of all three agencies in the previous period. So does the combined wisdom of all three tell me something about S&P’s next move? And then I repeated that for Fitch and Moody’s.
The first thing that strikes me, looking at the predictive models of S&P, Moody’s and Fitch is this: they all look similar at the top. High level, there is agreement between the three. The next thing that we see is that all three models single out the AAA (or Aaa) ratings first. Looking into the AAA ratings, you can see that they don’t change much, even when the outlook is negative. A negative outlooks basically serve as a warning sign, waving the red flag at the government. Obviously, having a AAA rating is a valuable asset and countries guard it with care to keep it as long as possible.
Interacting with these three models, you’ll find that most credit ratings don’t change much from one period to the next. Hovering over a decision node in the predictive models, you’ll see the distribution of the values for that node. It is in those details that you’ll find the exceptions. An interesting set of data and models to play around with.
If you want to create your own models with this (BigML format) dataset, you can clone it here in our Marketplace for free.
You can also take the data from the source, add any data to it that you like and create your own source file to upload to BigML and start modeling.