Explainer: What Is Model Drift?

[AnayDhawan]

Closes #36.

Concept: Model drift is the ongoing process by which a deployed model's accuracy and fairness gap change over calendar time as the population or the feature-outcome relationship it was trained on keeps moving.

Scoping against distribution-shift.md: This repo already has a distribution-shift explainer covering covariate/label/concept shift with one-shot reference-vs-current statistical tests (KS test, chi-squared). This explainer is deliberately scoped as the operational counterpart: rolling-window monitoring of a model already in production, using PSI and a Page-Hinkley change-point test, neither of which distribution-shift.md covers. The explainer states this boundary explicitly in its "Why It Matters" and "Related Concepts" sections so it doesn't read as a duplicate.

What's included:

• The German Credit Lending audit's age-based fairness gap, re-measured across 5 sequential windows of the same dataset instead of one random split. The gap ranges from 4.3% to 15.1% depending on which 200-row slice is measured, versus the audit's own single-snapshot 6.39% figure (not statistically significant at n=1,000, p=0.348).
• PSI computed between the first and last window on 3 features: credit_amount (0.119, moderate shift) actually moved more than age itself (0.096, borderline).
• Honest limitation Audit: German Credit Lending Bias #1: the dataset has no timestamp, so row order is used as a proxy for arrival order, and the explainer says so directly rather than implying this static academic file shows real calendar drift.
• Runnable detection code: population_stability_index, page_hinkley_test, rolling_fairness_gap.
• External case: Roberts et al. 2021 (Nature Machine Intelligence), a systematic review of ML COVID-19 diagnostic models (2,212 screened down to 61 meeting quality criteria, none found clinically usable), verified via search before citing rather than from memory.

All numbers were generated by actually running the code in the explainer against the repo's own German Credit Lending/credit_customers.csv, not invented.

[vercel]

@AnayDhawan is attempting to deploy a commit to the yashkewlani2020-gmailcom's projects Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Hey @AnayDhawan, your first PR on Fair Code, that's awesome, thank you.

This project is about making AI more accountable, and contributions like yours are what keep that work going. We'll review your changes shortly.

If you haven't already, give the contributing guide a quick read: it covers how audits are structured and what we look for in a review.