Customer Churn Prediction

2026-07-15In progress

An end-to-end classification project — from exploratory analysis and feature engineering to model comparison and an honest evaluation of what the model can and cannot predict.

classification scikit-learn pandas feature engineering

In progress — write-up coming

This write-up will be completed as the project is finished. The sections below show the problem and intended approach; results only get added once they’re real.

Problem

Churn prediction is the classic first “real” machine learning problem for a reason: the business question is concrete, the data is messy in realistic ways, and the interesting part isn't the model — it's deciding what “likely to leave” actually means, dealing with class imbalance, and being honest about what the model can and cannot tell you. This project is my end-to-end demonstration of that full workflow, built to be read by a hiring manager, not just scored on a leaderboard.

Data

[PLACEHOLDER: e.g. Telco Customer Churn] ([PLACEHOLDER: e.g. Kaggle / IBM sample datasets])

Approach

Start with exploratory analysis to understand class balance and feature distributions, then establish a simple baseline — majority-class prediction and logistic regression on raw features — before anything fancier. The evaluate step routes conditionally: iterate on features while a candidate model isn't clearly beating the baseline, move on to the write-up once it is.

Intended workflow: explore the data, engineer features, evaluate against a baseline with cross-validation, and iterate via error analysis until the model clearly adds value, then report honestly. — intended workflow — diagram will be updated if the approach changes during the build

From there the loop is deliberate: feature engineering informed by error analysis on the specific customers the model gets wrong, and model comparison with cross-validation rather than a single lucky split. Every experiment gets compared against the baseline, so the final write-up tells the honest story of what actually added value — and what didn't.

Result

Results will be published when the project is complete. No metrics appear here until they’re real.

demo video

a short recorded walkthrough goes here once the project is complete

What I'd do differently

Filled in honestly once the project is done — including what didn’t work and what I’d change with hindsight.

repo and notebook links will be added when the project is published