core practice · implementation task
Diagnose overfit with a train/holdout split
Fit a degree-1 vs degree-5 polynomial to 30 noisy points (50/50 split). Report train vs holdout MSE for both. Which overfits and how do you know?
The mechanism
Bias is systematic error from too-simple a model; variance is sensitivity to which sample you drew. Overfitting is low bias + high variance. The gym is holdout validation, not training R².
Expected evidence
Degree-5 lower train MSE, worse holdout MSE — classic overfit signature.
Try this drill in the practice workspace →
Hints when you need them
- Same random split for both models.
- MSE on holdout is the honest score.
Watch for these mistakes
- Tuning on the test set
- Chasing training loss without validation
- Equating complex with better