Churn Prediction Interpretably ML (GAM) +Retention

The Cost of
Friction

Linking Operational Failure to Financial Churn

"High sales with high churn is just 'burning through the market.' We transform abstract satisfaction into a hard financial metric: Retention Rate."

Decision Architect

↓

01. Transaction DNA

Integrating POS, Logistics, and Feedback into a single source of truth.

Revenue

POS

Azure SQL (Revenue)

  • Sales & Item mix
  • Order timestamps (Kitchen/Bump)
  • CheckID & OrderModeID resolution

Logistics

LAST-MILE

Delivery Metrics

  • Driver assignment & Pickup
  • Lat/Long coordinate validation
  • Order-to-Door (OTD) calculation

Feedback

VOC

Voice of Customer

  • Call Center complaints
  • App ratings & Aggregator reviews
  • Specific Taxonomy (Cold/Late/Wrong)

02. The Friction Taxonomy

Classifying negative events to isolate impact on retention.

Operational Friction

+

ExceedSoS

Orders taking > 30 minutes from tap to door. The primary driver of immediate churn.

WasCancelled

Hard friction. The ultimate bad experience where no product is delivered.

The "Silent Sufferer"

+

Hypothesis: These represent the highest risk. An order with objective failure (e.g., Late Delivery) but no formal complaint.

Impact Metric

Churn Probability: 2.4x Higher

The "Ideal Order" Flag

+

A binary gold-standard flag. An order is only "Ideal" if it has:
0 Complaints | 0 Speed Violations | 0 Cancellations | 0 Friction Events.

03. The Prediction Logic

Architecture: Generalized Additive Model (GAM)

graph LR %% Styles classDef active fill:#fff,stroke:#111,stroke-width:3px; E1[Events] --> F1[FRICTION
SCORING] F1 -->|Weighted Features| M1[LinearGAM
MODEL] M1 -->|Smooth Functions| P1[CHURN
PROBABILITY] P1 -->|Risk > 0.7| A1[ALERT
Store Operations] click F1 callNode click M1 callNode click P1 callNode class F1,M1,P1 active;
Data Pipeline
ML Inference
↑ Click nodes for Model Specs

04. Store Ranking Engine

Using Interpretable ML (LinearGAM) to grade store health.

The LinearGAM Model

Unlike black-box models, Generalized Additive Models (GAMs) allow us to plot the relationship between performance and retention as a curve. This allows Operations to see precisely where the ROI of a fix begins to plateau.

Training Window 6-Month Rolling
Retention Target 90-Day Reorder %
Validation Time-Series Split

Strategic Output: A-D Grading

A

High Predicted Retention. "Ideal" operation. Focus on scaling successful patterns.

D

Operational Chaos. High sales but 'burning' the market. Immediate intervention required.

Outcome

Predictive fix-list for 1,500+ stores.