Training data (with noise) Fitted model True pattern
Scatter plot showing polynomial regression on training data. Higher degree polynomials fit training data better but generalise worse.
Model complexity 3 degree-3 polynomial
Training error —
Test error —
Verdict —

The sweet spot is the degree where test error is lowest — the model that learned the pattern, not the noise. Above it: the curve contorts to memorise every training point. Below it: the model is too simple to capture the shape. This tension is the bias-variance tradeoff.