A subscription app loses customers every month and the retention team contacts users at random. You build a model that ranks users by churn risk, test it against a simple baseline, and hand the team a list they can act on.
A data scientist finds out what the data says and whether a model can help: framing the question, testing hypotheses, building and judging models. The value is correct reasoning about uncertainty, not the fanciest algorithm.
Goal: Learn Python, statistics and your first machine learning model from zero, and learn to evaluate it honestly.
Before you start
Beginner · 7 weeks
Week 1: Python basics
Variables, lists, dictionaries, functions · Running scripts and notebooks · Reading errors · Virtual environments
Practice: Write five small functions with tests.
Practice: Write five small functions with tests.
Common pitfalls
Think first, then reveal
What is overfitting?
Think first, then reveal
Why compare to a baseline?
Think first, then reveal
What is data leakage?
Think first, then reveal
Accuracy is 95%. Is that good?
Agree with retention what "churned" means, what action follows a prediction and how success will be measured.
Turning a vague business worry into a question a model can answer, with a metric that matches the decision.
Why this pick
A model that predicts churn well but triggers no useful action is wasted work.
What the role must know
Beginner mistake
Statistics
Modelling
Communication
Think first, then reveal
What is overfitting?
Think first, then reveal
Why compare to a baseline?
Think first, then reveal
What is data leakage?
Think first, then reveal
Accuracy is 95%. Is that good?
Think first, then reveal
How do you choose a validation split for time series data?
Think first, then reveal
How do you handle class imbalance?
Think first, then reveal
When would you not use ML?
Think first, then reveal
How do you check an LLM labeller?
Think first, then reveal
Why can a better offline metric fail online?
Think first, then reveal
How do you monitor a model when labels arrive weeks later?
Think first, then reveal
When does an LLM beat a trained model?
Think first, then reveal
How do you handle causal claims from observational data?
Build one, measure it and put it on your CV. Each has tools, setup commands and milestones.