Back to projects

Food Delivery Churn Prediction

Predict customer churn and prioritize retention outreach.

Role

Data Analyst & Data Scientist — portfolio simulation

Timeline

2026 · Portfolio simulation

Tools

Python - Exploratory Data Analysis - Scikit-Learn - Random Forest - OOF Threshold Tuning - Streamlit - Metabase

Metabase dashboard.

Explore the Metabase dashboard.

MetabaseInteractive dashboard

Business context.

A churn model must support a business decision, not only rank customers. In this portfolio simulation, I explored customer behavior, compared predictive models, and selected a decision threshold under a minimum precision requirement to help prioritize potential retention outreach.

Analysis approach.

  • Used Python and exploratory data analysis to understand customer characteristics and churn patterns. Negative feedback emerged as the strongest differentiator between the churn-risk and loyal-customer groups.
  • Compared candidate models with Scikit-Learn and selected Random Forest as the final model.
  • Tuned the decision threshold with out-of-fold predictions and a business constraint of churn precision ≥ 40%, selecting a threshold of 0.56.
  • Evaluated the final model on a holdout test set: churn recall 76.5%, churn precision 61.9%, and ROC-AUC 0.834.
  • Deployed an interactive Streamlit prediction app and provided a Metabase dashboard for customer-data exploration and analytical insight.

Conclusion.

The project connects customer analysis, predictive modeling, constrained threshold selection, and deployment. It demonstrates a retention-prioritization workflow, the model metrics do not establish that a live retention campaign reduced churn.

A model that supports a retention decision.

This project simulates the work of a Data Analyst and Data Scientist for an online food delivery business. The workflow connects customer EDA, model comparison, a business-constrained threshold, and tools people can use to inspect results and make predictions.

Model Evaluation and Decision Policy

76.5%Churn recall

Holdout test set.

61.9%Churn precision

Holdout test set.

0.834ROC-AUC

Holdout test set.

0.56Threshold

Selected with out-of-fold predictions.

≥ 40%Precision floor

Business constraint during threshold selection.

From customer feedback to a usable workflow.

Python EDA highlighted negative feedback as the clearest differentiator between risk groups. Random Forest then provided churn-risk predictions, while OOF threshold optimization translated a precision requirement into the final decision policy. Streamlit exposes interactive predictions, and Metabase supports exploration of customer insights.