title	LendingClub - Loan Outcome Prediction with Fairness Auditing

LendingClub - Loan Outcome Prediction with Fairness Auditing

TL;DR

This project predicts whether an approved LendingClub loan is likely to be Fully Paid or Charged Off, then audits whether model outcomes differ across ZIP-derived demographic proxy groups.

The notebook export (Fairness.md) is still the source of truth. The model card and the short report under reports/ summarize the evidence without copying the same story in three places.

What is in this repo

Artifact	Purpose
Data-Cleaning.md	Data loading, filtering, and feature preparation
EDA.md	Exploratory analysis of the loan data
Models.md	Baseline and random forest experiments
Fairness.md	Notebook export with the fairness metrics and mitigation outputs
Discussion.md	Results framing and tradeoffs
MODEL_CARD.md	Intended use, limitations, and metric snapshot
docs/DATA.md	Data source, layout, and privacy notes
reports/FAIRNESS_REPORT.md	Compact report generated from the committed evidence

Problem

Peer-to-peer loan investors care about default risk, but a model that only optimizes predictive performance can hide unequal behavior across borrower groups. This project keeps the task simple:

1. Predict whether an approved loan will be fully paid or charged off.
2. Prioritize high precision for loans predicted as fully paid.
3. Audit whether predictions differ across demographic proxy groups.
4. Document the fairness tradeoffs instead of presenting a single score.

Data

The original project used LendingClub funded-loan data from 2007 through Q3 2018. Declined-loan records were excluded because they had fewer useful fields and substantial missingness. The raw data is not committed to this repository.

Important constraints:

• The project focuses on completed loans only: Fully Paid and Charged Off.
• FICO scores were not available in the public dataset used here.
• Sensitive race labels were not available, so the fairness audit used a ZIP3 proxy linked with Census-derived racial proportions.
• The ZIP proxy is useful for auditing patterns, but it is not an individual race label and should be treated as a noisy proxy.

See docs/DATA.md for data access and expected local layout.

Method

The original workflow is notebook-first and includes:

• Data filtering and cleaning from LendingClub CSV exports.
• Feature preparation for tabular credit and loan attributes.
• Baseline decision tree experiments.
• Random forest modeling for the final investor-oriented ranking workflow.
• AIF360 fairness metrics for protected-proxy group comparison.
• Reweighing as a preprocessing mitigation strategy.

Evidence snapshot

The detailed values live in MODEL_CARD.md and are regenerated into reports/FAIRNESS_REPORT.md. The quickest read is:

Result	Value
Logistic regression accuracy	0.652441
Random forest precision	0.913286
Reweighed training mean outcome difference	0.000000
Reweighed test mean outcome difference	-0.002036

The random forest result should be interpreted with care: the original discussion optimizes for high precision and ranked loan selection, not broad deployment as a fully calibrated automated lending system.

Reproduce

The current repo is a rendered notebook archive. To reproduce the original work:

1. Download the LendingClub funded-loan CSV exports locally.
2. Place raw and intermediate files using the layout described in docs/DATA.md.
3. Run the notebooks in order from notebooks/: Data-Cleaning.ipynb, EDA.ipynb, Models.ipynb, Fairness.ipynb.
4. Regenerate the short report with python scripts/build_fairness_report.py.

The report script regenerates reports/FAIRNESS_REPORT.md from the committed metric snapshot and keeps the notebook evidence visible without repeating the same story in multiple files.

Responsible use

This is an educational and portfolio project. It should not be used for real credit decisions without a full compliance, legal, governance, and validation process. The ZIP3 proxy can expose group-level disparities, but it can also misrepresent individuals.

References

• Kamiran, F. and Calders, T. "Data preprocessing techniques for classification without discrimination." Knowledge and Information Systems, 2012.
• Bellamy, Rachel K.E., et al. "AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias." 2018.
• Emekter, Riza, et al. "Evaluating Credit Risk and Loan Performance in Online Peer-to-Peer Lending." Applied Economics, 2014.
• O'Neil, Cathy. Weapons of Math Destruction. Penguin Books, 2018.