Live · runs entirely in your browser
This is the fitted logistic-regression model from Churn Signal, exported to JavaScript with its exact preprocessing parameters. Change an account, test a contact policy, or score a CSV — no login and no server round trip.
Exported from models/churn_model.joblib · threshold 0.583 · holdout ROC-AUC 0.846
Holdout snapshot
Model liveThreshold metrics are replayed from the saved 1,409-row holdout evaluation. The scorer itself runs locally from exported coefficients.
01 · Single-account workbench
Every input updates the score immediately. The model uses 22 engineered features, including contract, service mix, spend, and tenure. The result is a prioritisation signal, not a causal verdict.
02 · Decision policy
A churn score becomes a campaign only after choosing a contact policy. The slider replays the saved holdout predictions; it does not refit the model.
Customers at or above this score are flagged for retention outreach.
0.583
selected policyThe selected policy maximizes F1 on out-of-fold training predictions, then is evaluated once on the holdout set.
Holdout trade-off
Precision is the share of contacted customers who churned. Recall is the share of churners who were contacted. Higher thresholds protect outreach capacity but miss more churners.
At 0.583, 270 of 374 churners are contacted while 222 retained customers are also contacted. That is a deliberate retention trade-off, not an accuracy claim.
03 · Batch scoring
Upload the raw Telco schema, validate every row in the browser, and download a ranked campaign file. Files are processed in memory on this device.
Required columns are the 19 raw feature columns; customerID is optional and
Churn is ignored if present. Missing IDs are generated, duplicate IDs are rejected,
and blank TotalCharges is filled from tenure × monthly charges.
No file loaded. Download the template or load the sample list to try it.
| Customer | Churn prob. | Decision | Risk tier | Tenure | Contract | Internet | Monthly | Annual at risk |
|---|
04 · What this is
The browser scorer is a deployment surface for the trained pipeline, not a new model and not a mock-up. The native Streamlit workbench remains the full reference implementation.
The logistic-regression intercept, one-hot coefficients, imputer medians, and scaler
parameters come directly from models/churn_model.joblib.
A Python export script regenerates the browser data so the static page cannot drift silently.
The browser derives the same AvgMonthlySpend, ServiceCount, and
IsNewCustomer fields, then applies median imputation, standard scaling, and
unknown-category handling.
Example profiles are checked against the Python pipeline's probabilities.
The threshold lab replays saved holdout predictions from the original evaluation split. It does not tune on the holdout set or claim production readiness.
Use the model card and deployment notes before applying scores to real customers.