CLV Intelligence

What it is

A portfolio analytics project demonstrating Customer Lifetime Value modelling across CRM transactions, GA4 web behaviour, and paid media data. The public demo runs on 6,538 synthetic customers — authenticated users can upload their own data and get a personalised report.

Live demo

The dashboard at clv-intelligence.onrender.com runs the full pipeline on synthetic data: 3.5 years of transactions, realistic channel mix, and correlated GA4 engagement signals.

How the model works

BG/NBD (Buy Till You Die) models each customer’s latent “alive/churned” state from their purchase history — frequency, recency, and time since first purchase. This is the part historic-spend metrics miss: a customer who bought once three years ago and one who bought once last month look identical in a simple RFM table, but have very different p(alive).

Gamma-Gamma then predicts expected order value from the distribution of each customer’s repeat transactions. Combining both gives a forward-looking CLV that accounts for both how often someone will buy and how much they’ll spend.

Segmentation

Four operational tiers using percentile-based thresholds (not k-means):

Segment	Threshold	Recommended action
High potential	Top 15%	Concierge service, max ad budget
Loyal	15–45%	Priority support, upsell opportunity
At risk	45–70%	Win-back email, soft discount
Low value	Bottom 30%	Self-serve, low-cost nurture

Thresholds are configurable in engine.py.

Data sources

The model degrades gracefully — CRM alone produces full CLV scores. Each additional source enriches it:

• CRM (required): transaction history — customer ID, date, order value
• GA4 (optional): aggregated session signals add an engagement score (+15% CLV lift max)
• Media spend (optional): unlocks the CLV:CAC matrix by channel
• Customer profiles (optional): demographic/firmographic enrichment

Column aliases are automatically resolved — revenue, amount, gmv, and total all map to order_value.

Authenticated upload

Users request access via a form (name, email, company). I get a Discord notification, manually add a passcode to allowlist.json, and reply with the code. The token is valid for 24 hours and scoped to their session’s uploaded files.

Stack

• Model: Python · lifetimes (BG/NBD + Gamma-Gamma) · scikit-learn
• API: FastAPI on Render · serves the React frontend as static files
• Frontend: React 18 · Vite · Tailwind CSS · ECharts (via echarts-for-react)
• Auth: manual OTP allowlist — no OAuth, no database