Every application scorecard is trained on a biased sample. To learn who pays back, a model needs outcomes, and a lender only observes outcomes for the applicants it approved. The people it rejected never got a loan, so they never had the chance to pay or default. The training data is whatever the previous credit policy let through.
That would not matter if the new model were only ever applied to the same kind of applicant. It is not. A scorecard scores everyone who applies — the full through-the-door population — including the profiles the old policy filtered out. And the decisions that move the business happen exactly there: around the cutoff, and in the segments you are considering approving for the first time.
Reject inference is the set of techniques that tries to correct this selection bias by estimating how the rejected applicants would have performed, and adding them to the training data.
Why ignoring the rejects is risky
Rejections are not random. Applicants were turned down for reasons that the old policy believed were related to risk: thin bureau files, high debt ratios, short employment, certain channels. If those reasons were partly right, the accepted-only sample under-represents the riskiest part of the applicant population, and a model trained on it tends to be too optimistic precisely in the region where approvals would expand.
The problem is easy to miss because it does not show up in validation. If the test set is also made of approved applicants, the model is graded on the same filtered population it learned from. Gini and KS can look excellent while the model has never seen the people it will be asked to approve next. This is one more reason why choosing the cutoff deserves as much care as building the score.
The main reject inference methods
Ignore the rejects. Train on accepted applicants only. It is simple and transparent, and when the old policy was close to random in the region you care about, it can be good enough. Its weakness is the bias described above.
Assign every reject as bad. Treat all rejected applicants as defaults. It corrects in the right direction but overshoots: it assumes the previous policy was never wrong, and it pushes the new model to reproduce the old one.
Extrapolation. Fit a preliminary model on the accepted applicants, score the rejects with it, and use those predicted probabilities as their outcome. Everything depends on how well that preliminary model generalizes to people unlike the ones it was trained on — which is the very problem being solved.
Parcelling. Sort all applicants into score bands using the preliminary score. Within each band, look at the observed bad rate of the accepted applicants and assign the rejects in that band to good or bad in a matching proportion, often inflated because rejects in the same band are usually riskier than accepts. Parcelling respects the risk ordering of the score while letting the bad rate vary band by band.
Fuzzy augmentation. Instead of deciding whether each reject is good or bad, duplicate every rejected applicant into two records — one good, one bad — weighted by the inferred probability of each outcome. No reject is forced into a hard label, and the uncertainty is carried into training through sample weights.
There are more sophisticated approaches — two-stage and bivariate models that jointly model the approval and the default, or semi-supervised learning — but parcelling and fuzzy augmentation remain the workhorses in practice because they are explainable and auditable.
An honest caveat
Reject inference does not create information. It replaces missing outcomes with assumptions, and its results depend on those assumptions. The academic evidence is mixed: several studies find little or no gain from the classical methods, and some argue that their assumptions are rarely made explicit. The right way to use reject inference is as a documented, testable decision — not as a box ticked by default.
Two sources of real information are worth more than any inference when they are available: the bureau performance of applicants you rejected but who were approved by other lenders, and a small, controlled test that approves a sample below the cutoff to observe how they actually behave.
A practical checklist
- Declare the rejected population explicitly. You need an approval status for every applicant, not just a dataset of loans.
- Check the preliminary score. Parcelling and fuzzy augmentation are only as good as the score used to place the rejects. Use a reasonable model, or an existing score you trust.
- Choose the bands with care. Quantile bands keep sizes even; equal-width bands follow the score scale. Set a minimum number of records per band so that no band's bad rate rests on a handful of loans.
- Calibrate and cap the probabilities. Calibration (Platt scaling or isotonic regression) and minimum/maximum caps keep extreme predicted probabilities from dominating the inferred outcomes.
- Decide how much weight the rejects get. The rejected population's weight controls how strongly the inference influences the final model.
- Sanity-check the inferred bad rate. It should be higher than the accepted applicants' bad rate, and consistent with what the business knows about the people it turned down.
- Never evaluate on inferred outcomes. Inferred records belong in training only. They must stay out of the test set and out of every cross-validation test fold, or the model ends up grading its own assumptions.
- Compare against the accepted-only model and document the choice. If a regulator asks why the model approves a segment the old policy rejected, the answer should be on paper. Explainability starts with the training sample.
How Thorcast handles it
In Thorcast, reject inference is an optional step between data preparation and modeling. It appears when the target is binary and a rejected population has been declared — an approval status column with its approved and rejected values — and those rejected records are still in the dataset.
The step supports parcelling and fuzzy augmentation. The preliminary score can be generated automatically or taken from an existing score column, and the advanced settings expose the number of bands, the banding method (quantiles or equal width), the minimum records per band, the rejected population weight, probability caps and calibration (none, Platt scaling or isotonic regression).
Before anything changes, a preview compares the original approved-only dataset with the augmented one: total records, observed bads, rejects inferred, inferred bad rate, total bad rate, a summary per band and any warnings. The user then decides whether to use the augmented dataset or skip the step. Inferred rows are used for training only: they never enter the test partition or a cross-validation test fold, so every reported metric is computed on real outcomes.
Reject inference will not turn a weak dataset into a strong model. Used deliberately, with its assumptions in plain view, it makes the scorecard you deploy closer to the population it will actually face.