Prediction intervals
summary
Covers concrete methods and code: training ensemble of quantile regressors (scikit‑learn, XGBoost, PyTorch), reducing probabilistic regression to binning + predict_proba cumulative sum with a scikit‑learn meta‑estimator, simple Bayesian Ridge and Gaussian Process baselines, and XGBoostLSS. Evaluates with pinball loss, CRPS, coverage and reliability diagrams; shows TabPFNv2 pretraining use and MAPIE/Conformal Quantile Regressors limitations. Includes synthetic and real examples (trip‑time, VaR), numeric coverage targets (e.g. 90% for 5/95 quantiles) and which approaches succeed or fail.
AI-written summary · about
chapters (11)
- 0:00Prediction intervals intro
- 2:33Probabilistic regression matters
- 6:43Quantile regression for intervals
- 11:18Regression objective functions
- 13:26Squared error in polynomial regression
- 16:36Triplet models for intervals
- 19:04Binning target as classification
- 20:51Evaluating interval models
- 23:39Reliability diagrams for calibration
- 25:24Failed heteroscedasticity methods
- 26:25Q&A
related talks
27mModelling the extreme using Quantile Regression
19mUnderstanding, Generating, and Evaluating Prediction Intervals
33mAdaptive Prediction Intervals
30mConformal Prediction: uncertainty quantification to humanise models
29mQuantile Regression Workflows
30mBoosting AI Reliability- Uncertainty Quantification with MAPIE
33mQuantifying uncertainty in machine learning models
01h 24mBoosting AI Reliability: Uncertainty Quantification with MAPIE