Quick answer: a credit-rating migration model treats ratings as states and estimates the probability that an obligor moves from one state to another over a chosen horizon. Those probabilities are arranged in a transition matrix. If the model assumes that the next state depends only on the current state, it behaves like a Markov chain. Default is often represented as an absorbing state: once the process enters default, it stays there for the purposes of that model horizon.

A rating is not a destiny. It is a state in a probabilistic transition system.

Why this is a mathematics problem

Rating migration combines conditional probability, matrices, repeated multiplication, absorbing states, survival, calibration and statistical estimation. It is one of the clearest examples of how a simple-looking table can encode a dynamic system.

It also reinforces a crucial probability lesson from A Probability of 70% Does Not Promise Seven Successes Out of the Next Ten. A 2% one-year default probability does not mean a borrower is “2% defaulted.” It means the model assigns a probability to a future state under specified conditions.

1. Build the state space

Suppose a simplified model uses three states: Investment Grade (I), Speculative Grade (S) and Default (D). A one-year transition matrix might be:

From / To	I	S	D
I	0.90	0.08	0.02
S	0.05	0.80	0.15
D	0.00	0.00	1.00

Every row sums to 1 because, at the end of the horizon, the obligor must be somewhere in the model’s state space. The default row is [0, 0, 1], making D an absorbing state.

2. Matrix powers turn one-year probabilities into multi-year probabilities

If the process is time-homogeneous and Markovian, the two-year transition matrix is P², the three-year matrix is P³, and so on. This is where linear algebra becomes a time machine.

Using the simple matrix above, an Investment Grade borrower can reach default within two years through several routes: default directly in Year 1, remain Investment Grade then default in Year 2, or migrate to Speculative Grade then default. The two-year probability of ending in default is:

0.90×0.02 + 0.08×0.15 + 0.02×1 = 0.05, or 5%.

That 5% is not obtained by simply doubling the one-year 2%. Pathways matter.

3. What the Markov assumption really says

A first-order Markov model assumes that, conditional on the current rating, the next transition does not need the entire path that came before. A firm rated BBB today is treated the same whether it arrived there from A last year or has been BBB for five years, unless the state definition is expanded to include that history.

This assumption is mathematically convenient. It may also be wrong. Momentum, recent downgrades, macroeconomic state, sector conditions and time already spent in a rating can all contain information beyond the current label.

4. Default as an absorbing state

In many credit-migration models, default is absorbing because the model’s job ends once default occurs. The process does not attempt to model restructuring, cure, recovery or a new post-default rating inside the same state space.

This is an example of model boundary rather than metaphysical truth. A company can emerge from restructuring in the real world. The absorbing-state assumption simply says that, for this particular transition problem, default terminates the rating path being measured.

5. Real rating matrices are larger

Actual migration studies often use many states such as AAA, AA, A, BBB, BB, B, CCC and Default. The Basel Committee’s public credit-risk-modelling review includes an illustrative historical migration matrix with exactly this structure. A highly rated obligor has a high probability of remaining near its current rating, while lower-rated states show larger downgrade and default frequencies.

See Basel Committee — Credit Risk Modelling: Current Practices and Applications.

6. Cohort estimation versus duration estimation

One way to estimate a transition matrix is the cohort method: take the set of firms in each rating at the start of a period and count where they are at the end. If 1,000 A-rated firms begin the year and 900 remain A, 70 become BBB, 20 become BB and 10 default, the row probabilities are estimated from those frequencies.

A duration-based or continuous-time method uses more information about when transitions occur and how long each obligor spends in each state. In a continuous-time Markov model, a generator matrix Q can be estimated and the transition matrix for horizon t written as P(t)=exp(Qt).

The BIS Irving Fisher Committee bulletin on rating migration matrices explains this continuous-time approach and its usefulness when some transitions are rare.

7. Why rare transitions are statistically difficult

Suppose no AAA borrower defaults in a small five-year sample. A naïve empirical matrix may assign a zero default probability. That does not prove default is impossible. It may only mean the event is too rare for the sample to observe.

This creates a recurring modelling problem: the most important tail transitions can be the ones with the least data. Smoothing, pooling, Bayesian methods, continuous-time estimation or external data may help, but each introduces assumptions that must be made explicit.

8. Transition matrices can feed mark-to-market credit models

Credit risk is not only default versus no default. A downgrade can reduce the market value of a bond or loan because investors demand a larger credit spread. A migration model can therefore map each possible future rating to a future value, weight those values by transition probabilities, and generate a distribution of credit-related portfolio outcomes.

Federal Reserve research on mark-to-market credit models describes CreditMetrics-style systems in which ratings follow a time-homogeneous Markov process and powers of a transition matrix are used to obtain cumulative default probabilities over longer horizons. See the Federal Reserve paper on granularity adjustment for mark-to-market credit risk models.

9. A simplified algorithmic pipeline

Define rating states. Decide the granularity and whether withdrawn/not-rated states are included.
Define observation rules. Specify the measurement horizon, rating source and treatment of multiple transitions inside a period.
Clean the history. Remove duplicates, align entity identities and distinguish mergers, withdrawals and genuine credit transitions.
Count exposure time or cohorts. Estimate how much observation each state contributes.
Estimate transition probabilities. Build a row-stochastic matrix whose rows sum to 1.
Handle rare transitions. Decide whether smoothing, pooling or continuous-time methods are justified.
Construct longer-horizon transitions. Use matrix powers or exp(Qt), depending on the model.
Map states to economic outcomes. Default loss, credit-spread changes or capital consequences may differ by destination state.
Stress the matrix. Increase downgrade/default probabilities or condition on adverse macroeconomic states.
Backtest. Compare predicted migration frequencies with realised transitions.
Check stability. Estimate matrices across time windows, sectors and cycles.
Update. Recalibrate when rating behaviour or portfolio composition changes materially.

10. The problem of withdrawn ratings

If an issuer disappears from the dataset because its debt matures, it is acquired, its rating is withdrawn or coverage stops, the modeller must decide how to treat that observation. Simply discarding withdrawals can bias transition estimates if withdrawal is related to credit quality.

This is a data-censoring problem. The matrix is only as meaningful as the rules used to decide who remains observable.

11. Time homogeneity is a strong assumption

If P is assumed constant, the same migration mechanism operates in expansion and recession. Real credit cycles do not behave so politely. Downgrades and defaults often cluster during stress, while upgrades become more common in recovery.

Alternatives include separate through-the-cycle and point-in-time systems, macro-conditioned transition models, hidden-state regimes or time-varying generators. The trade-off is familiar: more flexibility can improve realism but also increases estimation noise and model complexity.

12. Failure modes

Small-sample zeros. An unobserved transition is mistaken for an impossible transition.
Rating-history blindness. Recent downgrade momentum is lost under a strict first-order Markov assumption.
Cycle blindness. One long-run matrix is used as if recession and expansion were identical.
Selection bias. Withdrawn ratings or failed issuers disappear from the sample incorrectly.
State compression. Broad rating buckets hide materially different risks.
Agency mismatch. Transition behaviour from one rating methodology is applied to another.
Portfolio mismatch. Corporate-bond migration data are applied to mortgages, sovereigns or SMEs without justification.
Path dependence. Current rating alone is insufficient to describe future transition risk.

13. Diagnostics and falsifiers

Do rows sum exactly to 1 after cleaning and smoothing?
Are there suspicious zeros caused by low observation counts?
Does P² resemble observed two-year migration frequencies?
Do recently downgraded firms migrate differently from long-standing firms in the same state?
Does a recession-year matrix differ materially from the long-run matrix?
Are withdrawals concentrated in particular rating states?
Do sector-specific matrices tell a different story?
What happens if default is not treated as the only absorbing outcome?

Suppose someone claims, “The current rating contains all the information needed to predict the next rating.” A falsifier would be strong evidence that two firms with the same current rating but different recent rating histories have materially different transition frequencies. That would show the state definition is too small for the job.

14. Verification and update triggers

recalculate matrices using independent code;
compare cohort and duration-based estimates;
reconcile entity histories and rating dates to source records;
test one-year predictions against realised migrations;
compare matrix powers with directly observed multi-year transitions;
re-estimate after major recessions, rating-methodology changes or portfolio shifts;
inspect whether smoothing creates probabilities unsupported by economic logic;
separate genuine model failure from ordinary sampling variation.

Connections across the finance-and-banking algorithms lane

Credit-scoring algorithms — probability estimation at borrower level.
Bank capital models — where credit parameters become regulatory constraints.
Market-risk models — a different distributional problem driven by market prices rather than credit-state migration.
One Counterexample Can Be Enough — useful when challenging the Markov and time-homogeneity assumptions.

Research anchors

The deeper lesson

A transition matrix looks static, but it encodes motion. Matrix multiplication turns one-step movement into paths. Absorbing states turn certain destinations into endpoints. The most important intellectual move is to remember that the matrix is not the world: it is an estimate of how states changed under particular definitions, data and periods. Good modelling keeps returning to the observed transitions to see whether that map still deserves to be used.

Educational note: This article explains credit-risk mathematics and public modelling concepts. It is not a credit rating, financial advice, investment advice, or a recommendation regarding any borrower, bank or security.

Small Group Tutorials

How Credit-Rating Migration Models Work: Transition Matrices, Markov Chains and Default Absorption