31. Train Test Split in Machine Learning
31.4 What Goes Wrong Without a Split
Why. Skipping the split makes you celebrate a fake win.
What goes wrong
- You “ace” training accuracy and ship.
- New Diwali week orders look different; the model fails.
- Manager loses trust in data work.
How to notice
- If someone reports only one accuracy number with no test set — ask “on which rows?”
- If train score ≫ test score, you may be overfitting (Chapter 38 later) or leaking.
- Always log: n_train, n_test, random_state, stratify yes/no.
Ravindra Bagale's Tip
Notebook title madhe “TEST SCORE” alag section theva. Train score la star deu naka. Dhyan rakho!