Trained on all your data? Now you can't tell if it actually works.

Hold out data the model never sees. Train on one set, test on the other. That's how you know it works.

The Standard

80/20 split. 80% training, 20% testing. Sometimes 70/30 or 90/10 based on data size.

What Goes Wrong

Data leakage. Info from test set leaks into training. Always split FIRST, then normalize or engineer features.

Time series. Don't randomly split. Use past to predict future. Test on later data only.

Imbalanced classes. Use stratified splitting to maintain class proportions.

Better Approach

Cross-validation. Split into 5-10 folds. Train on some, test on others. Rotate through all. Better estimates, especially with small datasets.