Application to the TPM Optimization in Credit Decision Making

Statistical transition probability matrices (TPMs), which indicate the likelihood of obligor’s credit rating migrating from one state to another over a given time horizon, have been used in various credit decision-making applications. Empirical TPMs calculated from historical data generally do not satisfy desired properties. An alternative method is to formulate the problem into an optimization framework [165], i.e., to find an optimized TPM that, when projected into the future based on Markov and time-homogeneity assumptions, can minimize the discrepancy from empirical TPMs. The desired properties can be explicitly modeled as the constraints of the optimization problem.

This TPM optimization problem is high dimensional, non-convex, and nonseparable and is not effectively solved by nonlinear programming methods. It however can be well addressed by the proposed parameter adaptive DE algorithm where domain knowledge can be efficiently utilized to improve performance. In this chapter, we apply the proposed algorithm to this TPM optimization problem and compare its performance to a set of competitive algorithms.