Zenera Labs Engineering PortfolioInternally developed · Capability Assessment

Customer Churn Prediction & Revenue-Recovery Analytics

A leakage-safe XGBoost churn model on synthetic retail data that quantifies revenue-at-risk and ranks retention interventions by ROI.

Predictive MLCustomer AnalyticsRevenue Recovery

Built by priyanka · AI Engineering Associate

Request Something Similar Talk to Zenera Labs

01See It In Action

The system, up close

Real captures from the running system — the demo, screenshots and architecture come first, then the deep dive.

Demo

Walkthrough of the churn pipeline and the client-facing recovery analysis.

Churn recovery & analysis prediction demo

The demo walks through prediction, revenue-at-risk and intervention ranking.

02The Problem

Why this needed to be built

Meridian Retail Group saw a 14% drop in repeat-purchase rate over two quarters despite stable traffic — leadership suspected silent churn among mid-value customers who stop buying without complaining.

They needed to predict churn early enough for retention to work, quantify the revenue at risk, and know which intervention pays back — all explainable and auditable for an internal team.

03The Zenera Solution

What we built, and how it changes the workflow

We built an end-to-end, leakage-safe churn-analytics pipeline on a synthetic 6,000-customer dataset (fixed seed 42) that benchmarks XGBoost against a recency baseline, estimates revenue-at-risk, and ranks interventions by scenario ROI.

Thirteen monthly time-based snapshots avoid future leakage; RFM and support/marketing features are computed only from history. XGBoost (threshold 0.63, chosen on validation) beats the recency baseline on PR-AUC (0.374 vs 0.209) and F1 (0.418 vs 0.289).

Revenue at risk = P(churn) × 60-day value proxy, surfacing high-value high-risk customers. An observational matched-cohort causal analysis and a scenario ROI model (no invented campaign cost) rank Discount / Loyalty Nudge / Personal Outreach. All data is synthetic (seed 42) — a methodology demonstration, not real Meridian figures.

Core design principle

No feature uses information unavailable at the prediction date (per-feature leakage audit)
Time-based split, never random shuffle — reflects real train-on-past / deploy-on-present use
Reproducible end to end from a single master seed

04Key Capabilities

What it can do

A focused set of capabilities that define the system — no marketable filler, only what actually runs.

Leakage-safe feature engineering

RFM, marketing and support features bounded to history with an explicit KEEP/REMOVE audit.

Churn definition

Eligible active customers; churn = zero purchases in next 60 days (~9.6% prevalence at latest snapshot).

XGBoost vs recency baseline

Benchmarked on PR-AUC, ROC-AUC, precision, recall, F1, Brier, precision@k.

Revenue-at-risk

P(churn) × 60-day value proxy to prioritize high-value high-risk customers.

Causal / uplift analysis

Observational matched-cohort ATT with covariate balance checks (CI includes 0).

Scenario ROI

Ranks interventions by assumed capture rates and contact costs — no fabricated campaign spend.

Reproducibility

A single MASTER_SEED reproduces byte-identical CSVs.

05How It Works

The end-to-end workflow

The system processing path, in order.

1
Generate synthetic data
Four connected CSVs (customers, transactions, marketing, support) over 18 months, seed 42.
2
Audit & define churn
Data-quality audit and an empirically validated churn definition.
3
Time-based split
13 monthly snapshots → 60% train / 20% val / 20% test, no shuffle across time.
4
Engineer & model
Leakage-safe features; recency baseline + XGBoost; threshold picked on validation.
5
Evaluate
PR-AUC / ROC-AUC / precision / recall / @k on TEST, scored once.
6
Revenue & ROI
Revenue-at-risk and scenario ROI from the frozen pipeline; a client deck summarizes it.

06Technology

Stack & tech choices

Language & ML: Python, XGBoost, scikit-learn, pandas, NumPy

Explainability: SHAP (scaffolded)

Causal: Propensity-score matching (observational)

Notebooks: Jupyter (data audit, EDA, modeling, causal, ROI)

07Validation

Tested, verified and honestly scoped

Validation & tests

TEST metrics are from a real executed run (not invented); synthetic-data caveats are documented throughout.

XGBoost PR-AUC 0.374 vs recency baseline 0.209; F1 0.418 vs 0.289
Precision@k: top 10% of TEST = 0.446 precision / 0.354 recall
Churn prevalence ~9.6% at latest snapshot (matches the brief's ~9%)
Causal ATT 95% CI includes 0 — reported as observational, not proof of causation

Security & hardening

No real customer data used — fully synthetic
Referential integrity verified (0 orphaned rows)
Leakage audit explicitly documents excluded future-knowledge features

Honest limitations

Customer Churn Prediction & Revenue-Recovery Analytics is an internally developed engineering prototype. We state limitations rather than overclaim production readiness.