user@devops:~$ cat README.md
Conformal Prediction -- coverage with guarantees
# Description
Conformal Prediction (Vovk / Romano / Candes) implemented from scratch in Python + scikit-learn (no MAPIE). Part 1: Split Conformal Regression on California Housing (train 10,836 / cal 4,644 / test 5,160) with GradientBoosting and absolute-residual scores -- qhat(alpha=0.10)=0.790, empirical coverage 88.68% (target 90%), constant width 1.58. Part 2: multi-alpha calibration curve (0.50 to 0.02) with mean |cov-target| error of only 0.79 pp -- Vovk's +1 correction sustains the finite-sample guarantee. Part 3: Conformalized Quantile Regression (CQR) with GBR quantiles alpha/2 and 1-alpha/2; adaptive widths (mean 1.86) that level conditional coverage across yhat quintiles (~89-91% everywhere) versus Split's conditional failure (Q5 only 73.8%). Part 4: conformal classification on Wine -- LAC (1-p_y) and APS (cumulative mass) at 100% coverage with mean set size 1.13 (93% singletons). Part 5: 10-fold CV+/Jackknife+ vs Split on n=400 subsample -- 92.3% coverage without a calibration hold-out. Part 6: conditional-coverage dashboard and 5-method summary. 6 visualizations.
# Key features
$ Split Conformal Regression with +1 correction (Vovk): finite-sample qhat and marginal coverage ~1-alpha
$ Multi-alpha calibration curve: mean |cov-target| error = 0.79 pp on California Housing
$ CQR (Romano et al.): adaptive-width intervals that fix conditional coverage
$ Finding: Split CP Q1=97% vs Q5=74% conditional; CQR levels to ~89-91% across quintiles
$ Conformal classification LAC + APS on Wine: 100% coverage, mean size 1.13 (93% singletons)
$ CV+/Jackknife+ (Barber et al.): 92.3% coverage without wasting a calibration split
$ 6 visualizations: intervals, calibration, adaptive CQR, LAC/APS sets, CV+ vs Split, dashboard
# Gallery
# Technologies used
