Reject-Inference
Raymond A. Anderson · Oxford University Press eBooks · 2021
Rejects had not the opportunity to perform. Marginal Rejects are often cherry-picked based upon other data, or cheapened through down-sells, which distorts an Accepts-only model. Reject inference addresses resultant distortions but is contentious. (1) The basics—i) pointers—basic considerations; ii) missing at random, or not; iii) terminology—data manipulation, allocation, methodology; iv) characteristic analysis—for reject inference; v) swap-set analysis—proposed versus past; v) population flow diagram. (2) Intermediate models—especially ‘known Good/Bad’, which may use bureaux’s performance data. Others are Accept/Reject and Cashed/Uncashed. Possible formulae are provided for extrapolated performance assignments. (3) Inference smorgasbord—i) supplementation; ii) performance surrogates; iii) reject is Bad; iv) augmentation; v) weight of evidence (WoE) adjustments; vi) iterative reclassification; vii) extrapolation of accept performance. (4) Favoured technique—involving i) fuzzy-parcelling—record cloning and weight adjustments; ii) extrapolation—graphical setting of performance-adjustment parameters; iii) attribute-level adjustments—where needed; v) practicalities—variable names and coding, with an example.