Panel about probabilistic forecasting
summary
Panel argues that probabilistic forecasting works when it uses actual throughput history and proper assumptions: Monte Carlo on empirical throughput needs lots of data and stationarity to converge, and it suits portfolio-level predictions, not single bespoke projects. Practical use is daily reforecasting as new throughput points arrive, bootstrapping in a spreadsheet is enough to experiment, and the common failure is turning an 85% percentile into a hard deadline. Success stories show useful ranges of features shipped per time window.
AI-written summary · about
chapters (9)
- 0:00Panelist introductions
- 1:49Position statements
- 6:27Probabilistic forecasting methods
- 11:24Empirical distributions and assumptions
- 18:19Monte Carlo simulation
- 25:19Stability and system modeling
- 33:30Handling uncertainty and precedent
- 40:06Debate on stability and predictability
- 42:56Closing remarks
related talks
29mProbabilistic demand forecasting at scale
24mPlatform-First Forecasting: Engineering Scalable ML Systems
42mRobust Automated Forecasting in Python and R
33mKickstart Your Probabilistic Forecasting
32mLeveraging Conformal Prediction for Calibrated Probabilistic Time Series Forecasts
40mMulti-Series Forecasting at Scale with StatsForecast
54mTime Series forecasting with NumPyro
37mForecasting Time Series Data at scale with the TICK stack