Quantifying Uncertainty in Time Series Forecasting
summary
Claims that conformal prediction produces theoretically calibrated, distribution-free prediction intervals by estimating error distributions via time-series cross-validation (example uses four windows for a 12-step horizon) and then adding quantile bounds to point forecasts. Implementation lives in Nixtla’s ml forecast stack and generates 80% and 90% intervals; benchmarks report top-1% on the M5 dataset and ~24% better accuracy and ~400% speedup versus AWS Forecast. Tests show narrower intervals for seasonal series and wider for noisy ones.
AI-written summary · about
chapters (24)
- 0:00Quantifying uncertainty in time series
- 1:07Importance of time series forecasting
- 2:12Foundational time series models
- 3:57Tourism forecasting in Australia
- 4:59Introducing ml forecast library
- 7:47Feature engineering in ml forecast
- 9:11Scalability of ml forecast
- 10:56Limitations of point forecasts
- 13:50What are prediction intervals
- 15:15Multi-quantile loss problems
- 16:22Statistical methods challenges
- 17:25Conformal prediction solution
- 19:33Implementing conformal prediction
- 21:23Q&A on model comparison
- 22:55Comparing ml forecast libraries
- 25:28Q&A on conformal prediction
- 26:53Conformal prediction data size
- 28:17TensorFlow and framework support
- 30:05Handling distribution changes
- 31:27Regime-changing models
- 32:32Q&A on performance and outcomes
- 33:14Q&A on data requirements
- 34:16Multivariate forecasting with exogenous
- 35:18Q&A on handling outliers
related talks
31mUncertainty quantification: How much can you trust your machine learning model?
36mQuantifying uncertainty in Machine Learning predictions
19mUnderstanding, Generating, and Evaluating Prediction Intervals
32mLeveraging Conformal Prediction for Calibrated Probabilistic Time Series Forecasts
45mUncertainty estimation at scale with functime
22mConformal Inference with Tidymodels
30mConformal Prediction: uncertainty quantification to humanise models
35mUncertainty Unleashed