Multi-Period Migration Through Matrix Powers

Why Multi-Period Migration Requires More Than One-Step Probabilities

A one-period rating transition matrix describes the probability of moving from each current rating to each possible rating one period later. For a multi-year horizon, the question is how probabilities accumulate across every possible sequence of intermediate ratings. Matrix multiplication provides a systematic way to add those paths without listing them individually.

The key calculation is a matrix power when the same transition probabilities apply in every period. That calculation is useful only when its assumptions are understood: in particular, the rating process must be modeled as Markov and the one-period transition matrix must remain constant over the horizon. These methods are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Use row vectors for rating distributions. If \(\boldsymbol{\pi}_0\) gives the probabilities of being in each state at the starting date, then after one period the distribution is \(\boldsymbol{\pi}_1=\boldsymbol{\pi}_0T\). In a homogeneous model, the same matrix applies in each period, so after \(n\) periods \(\boldsymbol{\pi}_n=\boldsymbol{\pi}_0T^n\). The entry \((T^n)_{ij}\) is the probability of being in state \(j\) after \(n\) periods, conditional on starting in state \(i\).

Why Matrix Powers Add All Intermediate Paths

The two-period result follows from conditioning on the state after the first period. To be in state \(j\) after two periods, the process must first reach some intermediate state \(k\), then move from \(k\) to \(j\). The events associated with distinct intermediate states are mutually exclusive, so their probabilities add.

The second equality uses the Markov property: conditional on the current state \(k\), the next-state distribution does not depend on earlier states. The final equality is precisely the definition of matrix multiplication. Repeating the argument gives \(T^n\) for \(n\) periods. Equivalently, each entry of \(T^n\) is the sum, over all sequences of \(n-1\) intermediate states, of the product of the one-period probabilities along each path.

The row-sum result is also a useful calculation check. Since each row of \(T\) sums to 1, \(T\mathbf{1}=\mathbf{1}\), where \(\mathbf{1}\) is a column vector of ones. Therefore \(T^n\mathbf{1}=T^{n-1}(T\mathbf{1})=T^{n-1}\mathbf{1}=\mathbf{1}\). Small discrepancies can result from rounding displayed entries; calculations should use unrounded values until the final reporting step.

Propagating a Starting Portfolio Distribution

The matrix \(T^n\) gives a result conditional on a specific starting rating. A portfolio distribution is handled by weighting the rows by the starting probabilities. If \(\boldsymbol{\pi}_0\) is a row vector, then \(\boldsymbol{\pi}_n=\boldsymbol{\pi}_0T^n\). The entries of \(\boldsymbol{\pi}_n\) are the portfolio’s modeled proportions or probabilities in each state; they sum to 1 when the initial vector does.

When a Matrix Power Is—and Is Not—Appropriate

The matrix-power calculation rests on two distinct assumptions. The Markov assumption says that, conditional on the current rating, the next rating is independent of earlier ratings. The time-homogeneity assumption says that the transition probabilities do not change from one period to another. A model may satisfy one assumption but not the other.

• If the process is Markov and homogeneous, use \(T^n\).
• If it is Markov but transition probabilities vary by period, use the ordered product of the period-specific matrices.
• If current rating alone does not contain enough information to determine the next-state distribution, a rating-only Markov matrix may be inadequate without expanding the state description or changing the model.

For period-specific matrices \(T_1,T_2,\ldots,T_n\), with \(T_1\) applying first, the multi-period transition matrix is \(T_1T_2\cdots T_n\). The order matters because matrix multiplication is generally not commutative. Replacing these matrices by a power of one chosen matrix imposes time homogeneity rather than merely simplifying notation.

Common Mistakes and Model Limitations

A frequent exam trap is to calculate a multi-period probability by multiplying just one transition probability along a single path. That gives the probability of that path, not the probability of reaching the destination by all possible paths. For a two-period transition from \(i\) to \(j\), the required value is \(\sum_k T_{ik}T_{kj}\), not merely \(T_{ij}^2\), unless the model happens to make all other paths impossible.

Another common error is to confuse the row and column conventions. Under the convention used here, distributions are row vectors and update by multiplication on the right: \(\boldsymbol{\pi}_1=\boldsymbol{\pi}_0T\). A column-vector convention is also possible, but its matrix orientation must be defined consistently; mixing the two conventions can produce incorrect results even when the arithmetic is accurate.

Finally, a valid matrix power is not evidence that the model is realistic. Rating transitions may vary with the economic environment, and the current rating may not capture all relevant information about an obligor. The calculation describes the implications of the stated Markov and homogeneity assumptions; it does not establish that those assumptions hold. Estimation of transition probabilities and their uncertainty are separate modeling questions.