Summarising: The three different models have different bias-variance decompositions. The constant model has considerable bias (red point cloud not centered around ground truth), but it has low variance (small spread). Order 5 has minimal bias, but high variance. Quadratic seems good compromise.