Can we tell in advance whose credit rating will be cut?

A credit rating is a committee's opinion

Agencies such as S&P give companies a grade, the rating, that says how likely they are to pay their debts. The scale has 23 rungs, from AAA (the safest) to D (default). When the rating falls, debt becomes more expensive and contract clauses can be triggered.

Touch a rung. Each article in the thesis grouped the rungs in its own way.

Six ways to pick out the companies that will fall

They all receive the same figures for each company and learn, from past examples, to draw a border between those that fell and those that did not. What changes is the shape of the border they can draw.

Our illustration with 60 made-up companies and only two figures per company. The thesis models use dozens of variables at once.

Almost 28,000 ratings, six in ten from the United States

The first article gathered, from the S&P Capital IQ Pro database, 3,453 rated companies from 20 countries, from 2018 to 2024, and 51 variables for each: 43 financial figures (debt, cash, margins, coverage) and 8 business assessments (industry risk, country risk, governance).

Because the mix of ratings varies a lot from country to country, the author trained one model per country. Touch a country.

With more companies, the models pull apart

Each model tried to get right the class (A, B, C or D) of companies it had not seen. In most countries, almost all topped 90%. In the United States, with more companies and a broader mix, the neural network and the trees pulled away from the traditional models. Touch a country.

Average accuracy over 5 repetitions of 10-fold cross-validation (Table 2.6). A single decision tree did worse; the thesis does not publish its figures.

Now the question changes: will this company fall in the next 12 months?

The second article used 31,151 company observations (2017 to 2022) and 46 variables. Instead of guessing the rating, each model answers yes or no: will the company be downgraded or default in the following year? And there is one model for each of the six risk bands.

To mimic real use, the models learned from 2017 to 2021 and were tested on 2022, a year they had never seen.

Catching who falls without crying wolf

When few companies fall, “being right a lot” is misleading: whoever says nobody will fall is almost always right and is of no use. That is why the thesis measures two things at once: of those that fell, how many the alarm caught and of the alarms, how many were real. The F1 combines both into a score from 0 to 1.

Adjust how sensitive the alarm is.

XGBoost and Random Forest made almost no mistakes in any band

The two tree models came close to the maximum (1) in all six bands; XGBoost reached 1.000 at both ends of the scale. Logistic regression and SVM did reasonably in the good bands and stumbled in band BB and in the riskiest ones.

Faced with 2022, SVM lost the most

A model can do well on the data it learned from and fail the following year. Compare, band by band, the cross-validation result (●) with the 2022 test result (◆). XGBoost barely moved; SVM plunged in bands BB and CCC.

Management and governance weigh most in every band; liquidity gains weight with risk

To explain the predictions, the thesis used SHAP, a method that measures how much each variable pushed the model's answer. On average, the assessment of management and governance came first in all six bands. Right below it, in the safer bands comes the industry; from band BB down, liquidity, the cash at hand to pay the bills.

Approximate values, read from the charts in Figure 3.3 (mean absolute SHAP value, the seven most influential variables in each band).

Without the previous rating, accuracy drops, above all in the riskiest band

Results so close to perfection raise suspicion. The author redid the analysis in two ways: without the size variables (revenue, debt in absolute values) and without the rating history (the previous rating and the agency's last rating action).

Without size, nothing changed. Without the history, the F1 fell, slightly for XGBoost and sharply for Random Forest. In the riskiest band, Random Forest went from 0.93 to 0.25. Touch a band.