Distance Correlation: Detecting Nonlinear Dependence That Pearson Correlation Misses
Compare Pearson correlation with distance correlation on parabolas, circles, sine waves, crosses and independent clouds. Add noise, change the sample size, and see dCor stay above its permutation null band where Pearson reports nothing. A Blender-built plotting easel and linked-pair prop, with invariance, double-centring identity and Gaussian closed-form checks built in.
Model evaluation for a regression problem always comes down to comparing predicted values against actual ones, then compressing every point's error into a single headline number — but MAE, RMSE, R² and MAPE compress that same set of residuals in genuinely different ways. This simulator renders a live 3D bar chart of a synthetic predicted-vs-actual dataset: green columns for the true values, amber markers for the model's predictions, and red connectors sized to each point's residual. Dragging in noise spreads every prediction randomly and moves MAE and RMSE together, while dragging in outlier strength concentrates error into a single point and pulls RMSE sharply ahead of MAE — the exact mechanism behind "RMSE is outlier-sensitive, MAE is outlier-robust." A metric-focus switch recolors the bars by each metric's own per-point weighting so the difference is visible, not just numeric, and four live readouts (MAE, RMSE, R², MAPE) plus the RMSE/MAE ratio update every frame.
Compare Pearson correlation with distance correlation on parabolas, circles, sine waves, crosses and independent clouds. Add noise, change the sample size, and see dCor stay above its permutation null band where Pearson reports nothing. A Blender-built plotting easel and linked-pair prop, with invariance, double-centring identity and Gaussian closed-form checks built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install