Why this matters
Organisations need to plan before things happen. A lender wants to know which borrowers may struggle so a field officer can call early. An agritech company wants yield estimates before harvest to arrange storage and buyers. A health programme wants to know which months will be busiest at clinics. Predictive models turn past records into these estimates, and most data science adverts name machine learning or predictive modelling.
Prediction is not explanation
The statistics section asked "what caused this?" This section asks "what is likely to happen?" The two need different habits. A model can predict well using features that cause nothing: a household that tops up airtime often may be a good credit risk, but airtime does not cause repayment. That is fine for prediction, as long as you never present the model as proof of cause, and as long as the features are fair and lawful to use.
Models serve decisions
A good modelling project in this track starts from a decision, not an algorithm.
- What action will the prediction trigger? A phone call, a visit, a loan offer, extra stock.
- What does each mistake cost? Missing a struggling borrower costs one thing; calling a reliable one costs another.
- What will you compare against? A simple rule, such as "call everyone who was late last season", is the baseline the model must beat.
If a model cannot beat a sensible simple rule, it does not get used, however clever it looks.