A demand model for a bakery chain reports training RMSE 2.1 loaves and test RMSE 9.8 loaves. The model is a 12th-degree polynomial on daily temperature.
What is the diagnosis, and what is the sensible first move?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
A large gap between low training error and high test error is the signature of variance: the model has memorized the training days, including their noise, and that memory does not transfer. Lower the degree, add regularization, or feed it more data; the training error will rise a little and the test error will fall a lot.
Underfitting shows up as both errors being high and close together. A 2.1-versus-9.8 gap with a 12th-degree polynomial is the opposite problem.
Leakage makes the test score too good, not too bad. Here the test score is the disappointing one.
A small gap is normal. A test error nearly five times the training error means the model will disappoint in production.