The Cohort Estimator is a simple frequentist estimation of multi-state transitions. The estimator can be used to derive the transition probability matrix of a Markov Chain process with a finite number of states.
- Prerequisite for applying the approach is the definition of a suitable number of temporal cohorts and the allocation of all observed entities into such cohorts. The temporal intervals defining the cohorts are part of the design of the statistical analysis and depend on the domain.
- A second methodological aspect is the treatment of left/right censoring and transitions happening within a cohort interval. Multiple approaches are possible (first or last observed state within a cohort, average or longest living state etc.)
- Whether the different periods / cohort intervals can be consider homogeneous or not.
Single Period Cohort Estimator
In a single period cohort estimation all periods are considered equivalent (different samples of one-period transitions).
- The position in state space for an entity at discrete time is a random variable taking values in the state space .
- We assume a finite state space
- Cohorts are defined by equal temporal intervals
- Let be the number (count) of entities at state at time and at state at time .
- Let be the number (count) of all entities at state at time .
The directly estimated transition probability for that cohort is:
The estimator reflects that the probability of transition from m to n is the observed count number of all entities that migrated from m to n as a fraction of the count of all entities whose rating was m at k-1, that is, , irrespective of where they migrated to. The denominator includes all entities that did not migrate .
Under the assumption of homogeneity the estimates of different periods can be averaged. Using observation count as weights this results in an average transition matrix
NB: The count-weighted average will in general be different that the simple arithmetic average of estimated matrices over different periods.
Multi-Period Cohort Estimator
A multi-period estimator considers longer period intervals without the assumption of time-homogeneity. Computationally it is straight-forward extension:
where denotes the migration count of the period [k,l].
Confidence intervals for the cohort estimator can be estimated using the multinomial proportions method
Issues and Challenges
- State changes which occur within the period [k-1,k] are ignored. The time resolution (the number of cohorts) must be chosen so that there is as little ignored transition history as possible.
- The cohort estimator gives zero probability to a migration event that is not present in the data. E.g. in the presence of (right) censoring where we do not know what happens to the firm after the sample window closes (e.g. does it default right away or does it live on until the present) 
- Left truncation where firms only enter sample if they have either survived long enough or have received a rating.
- Simultaneous Confidence Intervals and Sample Size Determination for Multinomial Proportions Cristina P. Sison; Joseph Glaz, 1995
- Measurement and Estimation of Credit Migration Matrices by Til Schuermann Yusuf Jafry, 2003