Churn Prediction

Score which customers are likely to leave. Trains a binary classifier on historical customer features + an outcome label.

Quick start

1. Open Tools → Churn Prediction (/tools/churn).
2. Drop a CSV where each row is a customer and one column is the churn label (0/1, yes/no, etc).
3. (Optional) Specify the label column name.
4. Click Train & predict.

CSV requirements

• At least 30 rows.
• At least 5 churned and 5 retained customers.
• A column called churn, label, or target — or any column with only 0/1 / yes/no values, which RINK auto-detects.
• Numeric feature columns are used as-is.
• String columns with ≤ 20 unique values are one-hot encoded.
• High-cardinality strings (customer IDs, names) are ignored automatically.

How it works

RINK fits a RandomForestClassifier with 300 trees, max depth 8. The data is split 80/20 stratified by the label; the held-out 20% is used to compute accuracy, AUC, and the confusion matrix.

The trained model then scores every row in your dataset for the "top at-risk customers" list.

Reading the output

• Accuracy — fraction of test-set predictions that were correct.
• AUC — area under the ROC curve. 0.5 = random; 1.0 = perfect.
• Feature importance — which signals the model relies on most.
• Risk buckets — high (probability ≥ 0.70), medium (≥ 0.40), low.
• Top at-risk — 10 highest-probability customers with their feature values inline.

API

POST /api/churn/predict (multipart/form-data)

Field	Type	Default	Description
file	file	—	CSV, ≤ 10 MB
label	string	auto-detect	Name of the label column

Returns { accuracy, auc, feature_importance, risk_distribution, confusion, top_at_risk: [...] }.