Score contribution per author:
α: calibrated so average coauthorship-adjusted count equals average raw count
This is the first paper to propose a flexible local logit regression for defaulted loan recoveries that lie in [0,1]. Via a simulation study, we demonstrate that the proposed model is robust to nonlinearity, and non-normality of errors. Applied to Moody’s dataset, the local logit model uncovers the intrinsic nonlinear relationship between loan recoveries and covariates, which include loan/borrower characteristics and economic conditions. We exploit the empirical features of the local logit model to improve the specification of the standard regression for the fractional response variable (RFRV) model, which we refer to as the calibrated-RFRV model. The estimation of the calibrated-RFRV model is more straightforward and faster than the local logit model. The overall out-of-sample predictive performance of the calibrated-RFRV is superior to the local logit, RFRV, neural network (NN), regression tree (RT) and Inverse Gaussian (IG) models. The local logit model outperforms others in quantile forecasting, showing the attractiveness of this model for estimating tail risks, the accurate estimation of which is beneficial to risk managers.