A model's raw score, like 72 marks on a test, tells you little without context. The real question is whether that result is an improvement, sufficient, or just luck. This is the core of model evaluation: moving beyond predictions to ask if a model is trustworthy, generalizes to new data, and learns patterns rather than memorizing.

R-squared (R²) is the most common metric, measuring how much better a model is than a simple baseline guess. It explains the percentage of variance a model captures. However, a high R² can be misleading, often signaling overfitting—where the model performs well on training data but fails on new data, much like a student who memorizes answers but fails the real exam.

To diagnose these issues, developers examine residuals—the difference between actual and predicted values. Random residuals indicate a healthy model, while patterns suggest missing information. Standard Error quantifies prediction reliability, showing the average distance from truth. Ultimately, comparing training and test performance remains critical; a large gap reveals overfitting and underscores the need for robust validation beyond single metrics.