Ratings, Migration & Transition Matrices

Every model so far has produced a number — a score, a PD, a spread. The oldest and still most consequential output in credit is a letter. Ratings move trillions in capital, sit inside bond covenants and central-bank collateral rules, and — through transition matrices — feed directly into portfolio models. This module is about what those letters mean, how they move, and what to do with the matrix that describes their movement.

The big three, and how they differ

S&P, Moody’s, and Fitch cover the overwhelming majority of rated debt. Their scales look interchangeable (AAA/Aaa at the top, D/C at the bottom, the investment-grade line at BBB−/Baa3) but the philosophies differ in a way that occasionally matters:

• S&P rates the probability of default, full stop. A senior secured bond and a junior unsecured bond of the same issuer can carry similar issue ratings if the default likelihood is the same (severity is handled via notching).
• Moody’s rates expected loss — PD × LGD. Structure and collateral are baked into the letter itself, not an afterthought. For the same instrument, S&P answers “will it default?”, Moody’s answers “how much will you lose?” — our Module 1 decomposition, split across two business models.
• Fitch follows a PD-centric approach closer to S&P and functions as the tie-breaker in the many mandates that require two ratings.

Three structural facts matter more than methodology brochures. First, the issuer pays — the conflict of interest is real, was central to the 2008 structured-finance disaster, and is managed (not eliminated) by regulation and reputation. Second, ratings are opinions with committees behind them, deliberately slow, built from analyst judgment on exactly the toolkit of Module 3 — ratios plus qualitative overlay. Third, agencies publish their track record: default studies and transition matrices, which is what makes everything in the rest of this module possible.

Through-the-cycle vs. point-in-time

The single most useful distinction for interpreting any rating:

• A point-in-time (PIT) measure estimates default probability given current conditions — where we are in the cycle, today’s leverage, today’s equity price. KMV’s EDF from Module 6 is aggressively PIT; it moves daily.
• A through-the-cycle (TTC) measure asks how the borrower would fare across a full cycle, deliberately filtering out the cyclical component. Agency ratings are TTC by design: a BBB should mean roughly the same creditworthiness in a boom as in a recession.

Neither is “right” — they answer different questions. PIT is what you want for pricing and early warning; TTC is what you want for stable capital planning and for covenants that shouldn’t trigger on every wobble (imagine rating-linked collateral calls under a PIT regime: pure procyclicality, margin calls exactly at the bottom). The practical consequence: agency ratings lag by construction. The famous cases where a name held investment grade until days before failure are partly that design choice being honest — and partly genuine misses. Banks’ internal Basel models are typically mandated to sit somewhere on the PIT–TTC spectrum and must document where; the same portfolio can need both a PIT PD (for IFRS 9 provisioning — Module 9) and a TTC PD (for regulatory capital) simultaneously.

The transition matrix

The workhorse dataset of the ratings world: a square matrix where entry $(i, j)$ is the probability that a credit starting the year at rating $i$ ends it at rating $j$. One row per rating, an absorbing column for default. Reading one is a skill worth thirty seconds of practice:

• The diagonal is heavy — 85–92% for most grades. Ratings are sticky (TTC design showing up in the data).
• Off-diagonal mass hugs the diagonal — moves of one notch dominate; AAA-to-B is essentially unobserved in a single year.
• The default column is monotone and convex — a stylized long-run BBB defaults around 0.2% in a year, B around 4–5%, CCC north of 25%. Each step down the scale multiplies the PD rather than adding to it.
• The last row is trivial — default absorbs (in these matrices; real workouts and re-emergences are handled separately).

Multiply the matrix by itself and you get the 2-year matrix; the $N$-th power gives the $N$-year view. This is the Markov assumption — next year’s rating depends only on this year’s — and it’s what the widget below does live:

Try this experiment: start at BBB and push the horizon out. Watch the distribution spread — mass leaks both up (a few BBBs become A) and down, the default bucket compounds, and the “still BBB” probability decays toward the matrix’s long-run behavior. Then start at CCC and note how fast the story resolves: distressed credits don’t stay distressed — they cure or they die, mostly within a few years. That’s the falling hazard shape from Module 7, now visible in matrix form.

Building one: cohort vs. hazard

From a ratings history database, the classic cohort method is exactly what you’d guess: take all issuers rated $i$ on January 1, count where they stand on December 31, divide. Simple, transparent, and standard — but it wastes information (a downgrade in February followed by default in November counts only as “ended in D”) and it can’t see fast paths through intermediate states. The duration (hazard) method — Module 7’s intensity machinery with a rating dimension — estimates a generator matrix of transition intensities from the exact timing of every move, then exponentiates. It uses all the data, produces smoother rare-transition estimates, and handles censoring cleanly. The downloadable workbook implements the cohort arithmetic and the matrix-power analysis so you can trace every number.

CreditMetrics: migration becomes mark-to-market risk

Until 1997, credit risk management meant default risk: will they pay? J.P. Morgan’s CreditMetrics reframed the question for a portfolio that gets marked to market: what is the distribution of my portfolio’s value in one year, given that ratings migrate?

The recipe for a single bond:

1. Take the issuer’s current rating; its row of the transition matrix lists the possible end-states and their probabilities.
2. Revalue the bond in each end-state by discounting its remaining cash flows at that rating’s forward spread curve. A BBB→A upgrade is a small gain; BBB→BB is a real loss — no default required.
3. In the default state, apply a recovery assumption (drawn from a distribution, in the full version).
4. You now have a value distribution for the bond — expected value, variance, and crucially a fat left tail driven by downgrade-plus-default.

For a portfolio, the missing ingredient is co-movement: do my issuers migrate together? CreditMetrics answers with a device you’ll recognize: each issuer’s migration is driven by a latent standard normal “asset return” — thresholds carve the normal distribution into rating buckets so each single-name distribution matches its matrix row — and issuers’ latent variables are correlated, typically via equity correlations. That is Merton’s structural logic (Module 6) recycled as a copula before anyone used the word. When correlations rise, downgrades and defaults cluster, and the portfolio’s left tail fattens dramatically — the theme that owns Modules 10 and 11.

Where this connects

• Ratings are the institutional output of Module 3’s analysis; transition matrices are their empirical track record.
• The generator-matrix view of transitions is Module 7’s hazard machinery, state by state; Jarrow-Turnbull’s ratings-based extension made that formal.
• Regulators consume this material constantly: risk weights keyed to ratings, IFRS 9 staging driven by “significant increase in credit risk” (a migration statement), stress tests as conditional transition matrices — all next in Module 9.
• CreditMetrics’ correlated-latent-variable trick is the exact bridge to portfolio credit risk (Module 10) and copulas (Module 11).

Try it yourself

The Excel workbook below contains the stylized 8-state matrix from the widget with every row documented, a matrix-power sheet computing the 2, 5, and 10-year matrices (pure cell formulas — you can audit the Markov chain step by step), cumulative default curves by starting rating, and a cohort-method mini-example that builds a transition matrix from a toy ratings history. Swap in any published matrix and the whole workbook recomputes.

Get new posts by email

One email per new article. No spam, no upsells, unsubscribe anytime.