rejectkit is a Python library designed to address sample-selection bias in credit modeling, specifically the issue of reject inference. In credit risk modeling, only approved applicants have known outcomes, while the rejected applicants' outcomes remain unknown, leading to biased models when applied to the full applicant pool. rejectkit brings together eight classic reject inference techniques under a single API, enabling practitioners to correct for this bias and, crucially, to evaluate whether these corrections actually improve model performance on their own data.
The library implements methods from three main families of reject inference. The augmentation family includes techniques that manufacture labels for rejected applicants, such as assigning labels based on model scores, splitting individuals into weighted good and bad versions, encoding assumptions about reject risk through weights, or borrowing default rates from similar approved customers. Another set of methods avoids manufacturing labels and instead reweights the approved customers or incorporates econometric control functions. The third approach uses semi-supervised learning, applying pseudo-labels to high-confidence rejects and iteratively retraining the model. Each technique is grounded in different assumptions about the nature of the rejection process, and the library emphasizes that no single method is universally superior.
A distinctive feature of rejectkit is its benchmarking capability. The tool allows users to simulate the reject inference process by masking labels in datasets where all outcomes are known, then measuring how well each technique can recover the hidden labels. The primary metric, auc_recovery, quantifies the improvement over a naive model trained only on approved applicants, with scores ranging from zero (no improvement) to one (full recovery to the oracle model), and negative values indicating a detrimental effect.
rejectkit is available as a Python package on PyPI and GitHub. It is intended for credit risk modelers and data scientists working in financial services who need to assess and mitigate sample-selection bias in their predictive models.
In the Developer Tools space, rejectkit takes a focused approach. It focuses on correcting sample-selection bias in credit models by enabling reject inference and evaluation on applicant data. It is built as an open-source project for credit risk modelers. rejectkit is open source under the MIT license. It runs on the web and the command line.
It is developed by Han-co, and the product first shipped in 2026. Development happens publicly on GitHub with 2 commits in the last 90 days. Key capabilities include reject inference, bias correction, and python library. The interface is available in English, Japanese, and Korean.
Latest indexed changes and source events
Where do rejected applicants go? Reject inference and rejectkit verified by the PulseGate indexer
Other apps tracked under the same category.
