Adaptive Prediction Intervals
summary
Compares specific adaptive interval methods: split conformal with quantile regressors (Conformal Quantile Regression), weighted residual schemes (LCP), partitioning via decision trees and forests (LowCart / LowForest), and LSF grouping by prediction beams. Shows exact conditional coverage is impossible in finite samples so algorithms target local conditional coverage; benchmarks on public datasets (from ~1.5k to tens of thousands of samples) find tree‑based partitioning gives the best conditional Winkler and coverage. Notes modularity, inference cost, and practical tradeoffs.
AI-written summary · about
chapters (13)
- 0:00Adaptive prediction intervals intro
- 1:50Conformal prediction basics
- 4:01Limitations of constant-width intervals
- 6:09Why conditional coverage matters
- 7:15Adaptivity in prediction intervals
- 10:07Conformal quantile regression
- 11:12Locally adaptive conformal prediction
- 12:39LowCART and LowForest
- 15:34Locally sorted forest
- 18:20Metrics for interval evaluation
- 21:54Benchmarking results
- 26:34Implementation tips
- 27:24Q&A
related talks
22mConformal Inference with Tidymodels
19mUnderstanding, Generating, and Evaluating Prediction Intervals
34mPrediction intervals
30mConformal Prediction: uncertainty quantification to humanise models
31mUncertainty quantification: How much can you trust your machine learning model?
37mQuantifying Uncertainty in Time Series Forecasting
44mAccuracy Is Not Enough: Building Trustworthy AI with Conformal Prediction
33mQuantifying uncertainty in machine learning models