Residual Plots

A model can look great on a single score like RMSE and still be quietly “getting things wrong” in a systematic way. Residual plots are a simple visual check that helps you see those hidden problems by looking at the model’s errors directly.

What residual plots show
A residual is the difference between what you observed and what the model predicted: observed − predicted. A residual plot graphs these residuals against something informative—most commonly the model’s predicted values or an input feature. If the model is well-specified, residuals should look like random noise centered around zero, with no obvious structure.

Common patterns and what they mean
Residual plots are especially good at revealing assumption violations and model misspecification:

• Curved pattern: the relationship isn’t truly linear; you may need polynomial terms, interactions, or a different model.
• Fan/megaphone shape (spread increases with prediction): heteroscedasticity; errors grow with the target level (common in house prices).
• Clusters or stripes: missing categorical variables, segmentation effects, or unmodeled groups.
• Outliers: unusual points that can overly influence fitted parameters.
• Residuals vs time showing waves: autocorrelation in time series; errors aren’t independent.

Practical example
In a house-price model, a residual plot might show small errors for inexpensive homes but large positive residuals for expensive homes. That suggests the model systematically underpredicts high-end properties—often fixed by transforming the target (e.g., log price) or adding features like neighborhood quality.

Why this matters in ML
In AI/ML workflows, residual plots complement metrics by diagnosing bias, unstable error variance, and data issues. They guide feature engineering, target transformations, and model choice (linear regression vs tree-based models). In Python, they’re commonly made with statsmodels, scikit-learn (with custom plotting), and seaborn.