amaedaqureshi/sia-churn-dataset

SIA Retention Engine — Churn Training Dataset

Preprocessed, schema-validated version of the public IBM/Kaggle Telco Customer Churn dataset, prepared for training SIA's churn risk model (Phase 2 of the SIA Retention Engine execution plan).

Source

Real data: IBM Telco Customer Churn dataset (https://www.kaggle.com/datasets/blastchar/telco-customer-churn), ~7,043 customers, ~26-27% churn rate.

What's real and what's synthetic

This dataset was built for a telecom (prepaid-style) retention use case, but the public Telco dataset is a US postpaid dataset with no network- quality or recharge fields. To fit SIA's Subscriber schema without pretending fields exist that don't, four columns are synthetically generated and clearly marked as such below. They are drawn independently of the churn label (verified by a unit test) so they cannot inflate model performance -- they exist only so the pipeline's shape matches what SIA needs, not to help predict churn.

Column	Real or synthetic	Source
user_id	Real	customerID
avg_monthly_spend	Real	MonthlyCharges
active_plan	Real (relabeled)	Contract -> Flexi/Standard/Premium
churn	Real	Churn
data_usage_gb	Synthetic	Normal draw, mean/std set per InternetService tier for plausibility
signal_strength_score	Synthetic	Beta(5, 2) draw, range [0, 1]
days_since_last_recharge	Synthetic	Poisson(lambda=15) draw
support_tickets_open	Synthetic	Poisson(lambda=0.6) draw

Splits

Stratified 70/15/15 train/val/test on the churn label, seed=42.

Files

• train.csv, val.csv, test.csv — SIA-schema columns + churn label.

Intended use

Training and evaluating the churn risk classifier for SIA's Monitor agent (Phase 2). Not intended as a general-purpose telecom dataset -- the synthetic columns are placeholders for real network telemetry SIA would use in production.