Diego S. · Customer Segmentation & LTV Analytics

Portfolio · Case Study

Diego S.

Data Science Customer Segmentation Predictive Modeling Marketing Analytics

Education

Florida International University

BS · Computer Science

Florida International University

MBA · Marketing & E-Commerce

Pennsylvania State University

Master of Applied Statistics

Available now

Expert-Vetted

Top Rated

100%Job Success

Case Study

Customer Segmentation
& LTV Analytics

Bilingual DTC Brand

Health Supplements

E-Commerce

83K

Customers analyzed

4

Behavioral segments

2

Storefronts (EN + ES)

01 · 09

Project overview

Key project facts

Client

Bilingual DTC Health-Supplements Brand

Industry

E-Commerce · Health & Wellness

Engagement type

Data Science · Consulting

Platform

Databricks · PySpark

Scope

Segmentation · LTV · Migration

Technical stack

Databricks PySpark Python scikit-learn XGBoost KMeans statsmodels Delta Lake Shopify API Google Analytics API Tableau

Engagement summary

A customer analytics platform for a bilingual DTC health-supplements brand running separate English and Spanish Shopify storefronts (~83K combined customers). Order, customer, and web-analytics data from both storefronts were ingested through PySpark pipelines on Databricks into a medallion (Bronze/Silver/Gold) Delta Lake architecture, then unified into a single customer table keyed on email. Customers were segmented with RFM — recency, frequency, and monetary value — using KMeans on each dimension, producing four behavioral segments: Window-Shoppers, One-Timers, Emerging Loyalists, and Loyalists. A cohort-based XGBoost model then predicted which lifetime-value tier a customer would reach from only their first four months of activity, and a migration analysis tracked how cohorts moved between segments over time. The findings drove a segment-level marketing strategy spanning discounting, channel investment, and reactivation.

20% → 50%

Loyalist Concentration

A cohort's Loyalists (~20% of customers) placed over 50% of all orders

81%

One-Timer Stickiness

Share staying in segment — pinpointed the reactivation opportunity

$163

Highest-LTV Channel

Email had the highest per-customer LTV despite a small order share

02 · 09

Technical Design · Customer Analytics

Data Pipeline

Databricks · PySpark · Medallion Architecture

Sources

Shopify EN

English storefront · orders + customers

Shopify ES

Spanish storefront · orders + customers

Google Analytics

web traffic · channel attribution

Call Center

support + order data

Ingestion

PySpark Pipelines

on Databricks

Shopify API

orders_en/es · customers_en/es

GA Reporting API

sessions · sources

Scheduled refresh

batch

Storage

Bronze

raw ingested

Silver

cleaned + conformed

Gold

Delta Lake · analytics-ready

Features

Unified customer table

keyed on email · EN + ES merged

RFM features

recency · frequency · monetary

Order metrics

AOV · units/order · days-between

Cohort assignment

monthly acquisition cohorts

Scope

~83K customers

EN + ES combined

One-year window

13 monthly cohorts

Delta tables

gold layer

Bilingual

English + Spanish

Technical Approach

Two separate Shopify storefronts — English and Spanish — plus Google Analytics and call-center data were ingested through PySpark pipelines on Databricks into a medallion (Bronze/Silver/Gold) Delta Lake architecture. Both customer bases were unified into a single table keyed on email, with storefront kept as a dimension, then enriched with RFM features and monthly cohort assignments.

Key Decisions

Unify EN + ES on email — one segmentation model across both storefronts, with language/storefront retained as a dimension rather than split into two models.

Medallion architecture — Bronze/Silver/Gold Delta tables keep raw, cleaned, and analytics-ready data separate and reproducible.

Cohort assignment at feature time — every customer is tagged to a monthly acquisition cohort so retention and migration can be measured over time.

03 · 09

Technical Design · Customer Analytics

Segmentation & LTV Modeling

RFM · KMeans · XGBoost

1

Inputs

Recency · Frequency · Monetary

Per customer: days since last order, order count, total revenue since acquisition · computed on the unified customer table

per-customer RFM

2

Clustering

KMeans · R / F / M

Separate KMeans clustering on each RFM dimension produces a recency cluster, frequency cluster, and revenue cluster — each ranked low to high

KMeans × 3

3

Scoring

Overall Score → 4 Segments

The three cluster ranks sum into an overall RFM score, mapped to four behavioral segments

4 segments

Segments

Window-Shoppers One-Timers Emerging Loyalists Loyalists

4

Temporal Setup

First-4-Months Cohort Scoring

Each monthly cohort is scored on its first 4 months of activity; segments are re-scored through the full window to measure migration — a clean train/predict time separation

13 monthly cohorts

5

Prediction

XGBoost Multiclass

XGBoost (multi:softprob) with balanced class weights predicts which LTV tier a customer will reach from early behavior · held-out test set

multi:softprob

6

Outputs

Migration + Targeting

Segment migration matrices, discount-sensitivity by segment, and channel-by-segment analysis feed a segment-level marketing strategy

strategy inputs

Approach

Customers are segmented with classic RFM, but each dimension is clustered independently with KMeans rather than bucketed by fixed thresholds — so the cut points adapt to the actual distribution. Segments feed a cohort-based LTV model that predicts a customer's eventual value tier from only their first four months of behavior.

Key Decisions

KMeans per RFM dimension — data-driven cut points instead of arbitrary quantile thresholds; the clusters reflect how this customer base actually distributes.

Cohort time-separation — features come from the first 4 months, the target from the full lifetime. Early behavior predicts eventual tier without seeing the future.

Balanced class weights — the LTV tiers are imbalanced (few high-value customers), so class weighting keeps the model from collapsing to the majority tier.

04 · 09

Analysis findings

A loyal core drives the revenue

A small, loyal core drives most of the revenue — and each segment responds to discounts differently.

Across monthly cohorts, the Loyalist segment — roughly a fifth of customers — consistently placed more than half of all orders. The One-Timer segment, by contrast, is highly static: most never advance to a higher-value segment on their own. But One-Timers are also the most discount-sensitive segment, which makes targeted second-order offers the clearest lever for reactivation. A flash-sale analysis confirmed a loss-leader effect — a record repeat-customer revenue month, with roughly half the products discounted and half full-price.

Before

One blended customer base with no view of who drives value, how segments migrate, or which customers respond to discounts.

After

Four behavioral segments with measured revenue concentration, migration likelihood, and discount sensitivity — each mapped to a specific marketing action.

20% → 50%

Loyalist concentration

~20% of a cohort's customers placed over half of all orders

81%

One-Timer stickiness

Stay in segment — the reactivation target, and the most discount-responsive group

$380K

Flash-sale record month

Highest repeat-customer revenue month — validated a loss-leader discounting strategy

05 · 09

01 / 04Customer Segmentation & LTV · Output

RFM Segment Overview

The full base scored on recency, frequency, and monetary value and split into four behavioral segments, each profiled by AOV, lifetime value, units-per-order, and subscription rate.

What this shows

Four behavioral segments

Window-Shoppers, One-Timers, Emerging Loyalists, and Loyalists, sized by share of the customer base.

Value per segment

AOV, LTV, and units-per-order quantified for each segment, so spend can follow value.

Subscription penetration

Subscription rate climbs sharply from One-Timers toward Loyalists.

Income distribution

Each segment's income mix surfaced to sharpen targeting.

Databricks · PySpark · K-Means · XGBoostFigures illustrative

06 · 09

02 / 04Customer Segmentation & LTV · Output

Acquisition Source & Lifetime Value

Where customers come from paired with what each channel is actually worth: traffic mix alongside AOV, LTV, and units-per-order by source.

What this shows

Channel mix

Direct drives 53% of customers; paid channels are a deliberate minority.

Value by source

Email and Direct customers carry the highest LTV; Paid Social the lowest.

LTV, not just volume

Reframes acquisition around lifetime value rather than raw traffic share.

Budget implication

Pinpoints which channels deserve more spend per acquired customer.

GA4 · Shopify · DatabricksFigures illustrative

07 · 09

03 / 04Customer Segmentation & LTV · Output

Segment Economics & Cohort Migration

How the segments compare on RFM and revenue, and how a single monthly cohort migrates toward Loyalist status over a full year.

What this shows

RFM characteristics

Loyalists average 5.3 orders and $437 revenue at 96-day recency, vs One-Timers at 1.0 / $63 / 387.

Revenue concentration

Loyalists are a small slice of the cohort but the majority of its revenue.

Cohort migration

Order share shifts from One-Timer-heavy to roughly 97% Loyalist across twelve months.

Retention signal

Quantifies how quickly the base consolidates into repeat buyers.

Databricks · PySpark · cohort analysisFigures illustrative

08 · 09

04 / 04Customer Segmentation & LTV · Output

Segment Migration Dynamics

Transition probabilities between segments a year on, showing which customers advance, hold, or slip back, and where intervention pays off.

What this shows

One-Timers are sticky

81% remain One-Timers a year later; only 12% advance to Emerging Loyalist, 7% to Loyalist.

Emerging Loyalists split

59% hold, 17% advance to Loyalist, and 24% slip back to One-Timer.

Loyalist retention

55% stay Loyalist; 44% soften to Emerging and only 1% lapse fully.

Highest-leverage nudge

The One-Timer to Emerging jump is where targeted offers move the most value.

Databricks · PySpark · transition modelingFigures illustrative

09 · 09