A hidden demand curve — price on the x-axis, weekly units sold on the y-axis — seen only through noisy weekly observations. Fit a polynomial of growing complexity and watch training error fall forever while test error turns back up. The gap between the two curves is the noise your model memorized.
The model: units = β₀ + β₁·price + β₂·price² + … + β_d·price^d, fit by least squares on the training weeks only. The test weeks never touch the fit — that is why their error is the only number the business should believe.