The Internal Rating Based (IRB) approach requires that financialinstitutions estimate the Loss Given Default (LGD) parameter not only based onclosed defaults but also considering partial recoveries from incomplete workouts.This is one of the key issues in preparing bias-free samples, as there is aneed to estimate the remaining part of the recovery for incomplete defaultsbefore including them in the modeling process. In this paper, a new approachis proposed, where parametric and non-parametric methods are presented toestimate the remaining part of the recovery for incomplete defaults, in pre-defined intervals concerning sample selection bias. Additionally it is shown thatrecoveries are driven by different set of characteristics when default is aging.As an example, a study of major Polish bank is presented, where regressiontree outperforms other methods in the secured products segment, and fractionalregression provides the best results for non-secured ones.