A rating transition matrix lists the probabilities of moving from each starting credit-quality category to every possible destination over a specified horizon. Its rows let a risk manager forecast upgrades, downgrades and default together, rather than treating every surviving borrower as having unchanged credit quality.
The categories may be agency credit ratings or an internal rating system. Always read the table's scope and horizon: a one-year issuer-rating table and a quarterly instrument-rating table describe different experiments.
Read across a starting-state row
Use the row convention
where one step is one year. Consider this fictional model, whose invented states are Good, Weak and Default:
| Starting state → state one year later | Good | Weak | Default |
|---|---|---|---|
| Good | 90% | 8% | 2% |
| Weak | 10% | 85% | 5% |
| Default | 0% | 0% | 100% |
The Good row says that a borrower starting Good has an 8% chance of being Weak next year. The reverse move is read from another row: Weak → Good has probability 10%. The diagonal records staying in the same category. Columns need not sum to one; each row must contain nonnegative probabilities summing to one.
Here Default is an absorbing state: its row is . It records that default has occurred, even if the real business later restructures. A published table showing only initially performing categories needs an explicit Default row before it can be used as a square transition kernel on this state space.
Build the denominator before estimating the entries
For a simple cohort estimate with complete observations, let be the number of entities starting in category , and the number ending in category at the chosen horizon. Then
The denominator is the starting row's population, not the whole sample or the destination category's population. Small starting cohorts make rare transitions particularly uncertain. A zero observed count is not proof that a transition is impossible.
A “withdrawn” or “no longer rated” column needs a stated treatment. Withdrawal is not automatically default or survival: an entity may cease to be rated for reasons such as acquisition. Deleting that column and rescaling the remaining entries changes the conditioning. Interpreting the rescaled row as the original population's future probabilities requires an assumption about the informativeness of withdrawal. McNeil, Frey and Embrechts discuss this explicitly in §10.2.1 of Quantitative Risk Management; CreditMetrics §2.2 also identifies its adjustment for no-longer-rated observations.
Two years means matrix multiplication
To turn the stipulated annual matrix into a two-year forecast, assume a time-homogeneous Markov chain: today's state contains the relevant predictive history, and the same annual probabilities apply in both years. An empirical one-year table alone does not establish either assumption.
The Chapman–Kolmogorov equations combine every possible intermediate state:
For example, Good → Default over two years is
The final term retains borrowers who defaulted in the first year. Because Default is absorbing, 4.2% is the probability of default by year 2, not the probability of a new default during year 2. New defaults during the second year account for of the original Good population. See credit migration for the distinction between cumulative and survival-conditioned default risk.
The complete two-year matrix is:
| Starting state → state two years later | Good | Weak | Default |
|---|---|---|---|
| Good | 81.8% | 14% | 4.2% |
| Weak | 17.5% | 73.05% | 9.45% |
| Default | 0% | 0% | 100% |
This is a matrix product, not an entry-by-entry square. For example, the two-year probability of finishing Good includes Good → Weak → Good, as well as Good → Good → Good.
A book is a mixture of starting rows
Let describe a randomly chosen borrower's starting state: 80% Good and 20% Weak. With row vectors, forecast by multiplying on the right:
Fictional annual model. Weight the two starting-state rows by their initial shares; all displayed percentages describe borrower counts, not debt values or loss amounts.
Thus the book's two-year cumulative default probability is . This is an expected fraction across borrowers with the stated composition, not a claim that exactly 5.25% will default. The matrix does not specify dependence between different borrowers, and counting borrowers equally does not weight their exposures.
For the underlying state and row conventions, Markov Chains — Transition Matrices and the Markov Property develops the model through transition diagrams and credit-rating states. Its public page is a preview; the full lesson requires membership.
What the matrix does not settle
Historical transition frequencies estimate real-world migration experience for a particular population and period. They are not automatically risk-neutral probabilities suitable for pricing traded claims. The numbers above are stipulated teaching inputs, with no rating-agency calibration.
If annual transition laws change over time, use the successive kernels in date order, , rather than the square of a single matrix. Rating history, economic conditions and the treatment of withdrawn entities can all matter. Finally, migration probabilities do not supply a bond's value in each surviving state, exposure at default or recovery: those are additional inputs to a credit-loss or valuation model.
Sources
- Gupton, Finger and Bhatia, CreditMetrics — Technical Document (1997), §2.2, printed pp.24–25; §2.3, pp.26–27: row interpretation, historical estimation uncertainty, no-longer-rated adjustment and the separate valuation step.
- Susan Holmes, Stanford University, Stats 366 / Stats 166: Markov Chains: matrix powers and row-vector distribution propagation.
- Alexander J. McNeil, Rüdiger Frey and Paul Embrechts, Quantitative Risk Management: Concepts, Techniques and Tools, revised edition (Princeton University Press, 2015), §10.2.1, printed pp.374–376: rating migration, withdrawals and business-cycle limitations.
- Richard Durrett, Essentials of Stochastic Processes, third edition (Springer, 2016), §§1.1–1.2, printed pp.1–2 and 9–11: homogeneous Markov transitions and multistep probabilities.