Risk Model Drift & Retraining Monitor — 2D

Market Regime

Regime drift speed 1.0×

Drift magnitude 0.35

PSI alert threshold 0.25

Auto-retrain when PSI crosses threshold

Live Monitoring

PSI (vs. baseline)

0.000

Simulated day

Model accuracy

96.0%

Days since retrain

Drift status

Stable — no significant shift

How it works

Every deployed credit/fraud risk model is scored against a live PSI check, comparing the distribution of scores it produces today against the distribution it was trained on:

PSI = Σᵢ (Aᵢ − Eᵢ) · ln(Aᵢ / Eᵢ)

Aᵢ = % of live scores in bin i (Actual)
Eᵢ = % of baseline scores in bin i (Expected)

Industry rule of thumb (used across Basel-III-regulated credit scoring):

PSI < 0.10 — no significant population shift, model stays in production.
0.10 ≤ PSI < 0.25 — moderate shift, investigate and monitor closely.
PSI ≥ 0.25 — significant drift; the scored population no longer resembles training data and the model should be retrained.

In this simulator the live score distribution executes a mean-reverting random walk whose speed and shock size you control, mimicking a changing macroeconomic/credit regime. This 2D edition draws the same rolling window of daily histograms as a ridge (joyplot): each receding blue row is one simulated day's distribution, drawn behind the previous day; the dashed grey ridge at the back is the frozen training baseline. Drag the canvas to pan the ridge and scroll/pinch to zoom into the most recent rows. Accuracy is modeled as degrading roughly linearly with PSI — the further live data drifts from what the model learned, the less reliable its predictions become. Retraining refits the model on the current population, instantly resetting PSI toward zero and accuracy toward its ceiling — exactly the "detect drift → retrain" loop described in real MLOps pipelines for financial risk models.

drag to pan · scroll/pinch to zoom

⚠ DRIFT ALERT

Live score distribution (by day) Training baseline (frozen)