Home▸Data Science▸Heteroscedastic Regression: Why Textbook Standard Errors Lie and Weighted Least Squares Wins

Heteroscedastic Regression: Why Textbook Standard Errors Lie and Weighted Least Squares Wins

Noise that grows with the predictor draws a fan of points. See OLS and weighted least squares both hit the true line, but compare the real standard error, the textbook one and the sandwich estimate, and watch 400 repeated samples show WLS slopes scatter less. The 3D depth axis is the standardised residual, with Monte Carlo and Gauss-Markov checks built in.

Data Science3DModerate60 FPS📱 Mobile-adapted🟠 Blender⇄ 2D version
Heteroscedastic Regression: Why Textbook Standard Errors Lie and Weighted Least Squares Wins ↗ Open standalone

Before any model-selection or hyperparameter search runs, an AutoML pipeline has to turn raw relational tables into a flat feature matrix — and it does this automatically with Deep Feature Synthesis. This simulator builds a small relational schema (Customers → Transactions → Categories), lets automated aggregation primitives (COUNT, SUM, MEAN, MAX, STD) walk the relationship graph at one or two hops of depth, and scores every synthesized feature against a hidden target with a live Pearson correlation. Watch the 3D scene light up each customer hub in proportion to how predictive its best auto-generated feature turned out to be, tune how much real signal exists versus noise, and see automated feature engineering either converge on the useful aggregate or, when the signal is weak, find nothing better than chance.

⚙ Under the hood

Noise that grows with the predictor draws a fan of points. See OLS and weighted least squares both hit the true line, but compare the real standard error, the textbook one and the sandwich estimate, and watch 400 repeated samples show WLS slopes scatter less. The 3D depth axis is the standardised residual, with Monte Carlo and Gauss-Markov checks built in.

AutoMLfeature engineeringdeep feature synthesisrelational dataPearson correlationThree.jsHeteroscedasticityWeighted Least SquaresSandwich Standard ErrorsGauss-MarkovRegression

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install

What did you find?

Add reproduction steps (optional)