CHURN RADAR_

Explore 7,043 customer records from IBM's public Telco sample. Every tile is computed in your browser. Select a bar to filter the view.

Interactive recreation of the private system, replaying recorded fixtures.

Filters None selected. Choose a bar below to slice the data.
Cohort compare
Loading cohort fixtures…
customers in view
churn rate in view
monthly $ lost to churn
Monthly charges from customers who churned.
predicted monthly $ at risk
Churn probability multiplied by monthly charges for this view.
model AUC (held-out)
Logistic regression outperformed gradient boosting.

Churn rate by contract

By internet service

By payment method

By tenure (months)

New customers churn most often. Loyalty builds over time.

By monthly charges

Expensive plans without long contracts carry more risk.

Small slicers

tech support
paperless billing
senior citizen

Retention over time: Kaplan-Meier survival by contract

Tenure is treated as time to event. Active customers are censored at their current tenure because this dataset is a snapshot. Curves respect every filter except contract itself.

Intervention simulator: from probabilities to money

Expected net equals save chance × Σ pᵢ × horizon × chargesᵢ for targeted customers, minus offer cost × targeted customers. The validation workbench shows held-out reliability. Move each assumption to test sensitivity.

expected net over horizon
customers targeted
expected saves
total offer cost
held-out precision at τ
held-out recall at τ

Loading held-out evaluation provenance…

The dot marks a unique optimum when one exists.

Historical backtest: held-out evaluation only

Realized outcomes at 45% threshold

Loading eligible held-out predictions…

Evaluation cohort and split provenance load from provenance.json.

Active-customer queue: outreach candidates only

Loading active customers…

IBM sample customer IDs; recorded churners are excluded. The table previews 50 rows; export includes the full filtered queue.

Validation workbench: 25% stratified hold-out

Loading calibration and audit evidence…

Reliability curve

loading bins

Diagonal = perfect calibration. Points are equal-width probability bins on untouched predictions.

Held-out segment audit

What-if scorer: the customer

The prediction

marker = dataset base rate

Why: top signals for this customer

Bars are each input's contribution to the log-odds: gold pushes toward churn, blue pulls away. This is the model's own arithmetic, not a post-hoc explainer.

CSV batch scoring: IBM Telco schema

Required: the 18 model inputs from the IBM Telco CSV. customerID and non-model columns may be included.

No file loaded. Choose a CSV or load sample rows.

Model card: held-out results

The logistic regression coefficients ship in model.json and run in the browser. The model was trained on the public IBM Telco dataset with a 25% stratified hold-out. TotalCharges was omitted because it overlaps with tenure multiplied by monthly charges.

0.00.20.40.60.81.0 0.2.4.6.81.0 logistic regression / AUC 0.845 gradient boosting / AUC 0.830 false positive rate true positive rate
Contract: Two year InternetService: Fiber optic tenure Contract: One year PhoneService: Yes StreamingMovies: Yes PaperlessBilling: Yes MultipleLines: Yes StreamingTV: Yes PaymentMethod: Electronic check -1.2501.25