With bank loans, both loss given default (LGD) and time to recovery following a default may vary significantly based on many factors, including, but not limited to, the obligor’s characteristics, collateral types, product features, industry sectors and macroeconomic environment. Since time to recovery is often right censored due to the LGD sample end-date cutoff, and if the LGD is correlated with time to recovery, the observed LGD may suffer from the censoring bias. A traditional LGD model (that is, one that omits time to recovery and ignores the censoring) will be biased when applied to nondefaulted performing loans in which time to recovery is unknown. To address this issue, a new modeling approach is proposed; it entails predicting both the LGD and the probability distribution of time to recovery. This is then illustrated using an existing workout LGD data set consisting of both censored and noncensored recoveries. Five model specification tests are performed. When compared with the traditional approach, the new approach is shown to fit the data better, resulting in a higher LGD prediction and marginal sensitivity to changes in the macroeconomic environment in stress testing.