Estimating future transition probabilities when the value of side information decays, with applications to credit modeling

ABSTRACT A number of conditional transition-probability models make use of side information: explanatory variable values known only initially that are useful for predicting transitions of the variables of interest. For example, the Cox proportional hazard model is used to provide future hazard arrival probabilities based on initially available side information. In this paper, we observe that a number of commonly used models, including the Cox proportional hazard model, force the side information value (which we define precisely) to persist over all time horizons, in contradiction to evidence drawn from historical data indicating that the value of side information exhibits pronounced decay over time. There is a solid intuition behind this decay: explanatory variable values typically evolve over time, so the incremental predictive value of the initially available side information should decline over time. Using information-theoretic methods, we introduce new, tractable and robust methods to estimate conditional transition-probability models that enforce decay in information value. That is, under our models, the impact of the explanatory variables on the conditional probability distribution must decay over time. We benchmark our methods against other methods on Compustat default transition data and Standard&Poor's rating-transition data. Here, the side information consists of initial observations of financial ratios and other explanatory variables. We find that our methods outperform the benchmark methods out-of-sample.