Prediction error

After fitting, look at the residuals $r_i=y_i-{\hat{y}}_i$. Ideally they look like random noise: centred on zero, with constant spread and no pattern.

Plot residuals against predictions or each feature. A curve means the relationship is non-linear — add polynomial terms or transform a feature. A funnel (spread growing with $\hat{y}$) means non-constant variance — try predicting $\log y$. Clusters suggest a missing categorical feature.

Points with huge residuals are outliers; points with extreme feature values are high-leverage points. A point that is both can strongly change the fitted line. Check them — they may be data errors.