Part Overview
A point forecast that is silent about its own uncertainty is dangerous, and a model that assumes the world stops changing is wrong the moment it ships. Part V addresses both failures. Its three chapters move from saying how sure a forecast is, to keeping a model accurate as data streams in, to building systems that detect their own staleness and adapt. This is the part that turns a good model into a trustworthy one, and it sits deliberately between the modeling parts and the decision-making part, because decisions are only as sound as the uncertainty they rest on.
Chapter 19 develops probabilistic forecasting and uncertainty quantification: the difference between point and probabilistic forecasts, proper scoring rules and calibration, quantile and distributional methods, Bayesian and ensemble approaches, and conformal prediction, which gives distribution-free coverage guarantees even for time series. Chapter 20 turns to online and continual learning: streaming optimization, concept drift and its detection, incremental models, and continual learning that resists catastrophic forgetting. Chapter 21 closes the loop with adaptive systems: retraining triggers, production drift detection, active learning, and human-in-the-loop adaptation, the operational habits that keep a temporal model alive.
The industrial sensor running dataset leads this part: telemetry drifts, and Chapters 20 and 21 use it to make drift concrete, connecting back to the change-point detection of Chapter 8. Chapter 19 closes with decision-aware evaluation, an explicit bridge to the next part, since a forecast that ignores how it will be used is being graded on the wrong test. Part V is short, but it carries the standards (calibration, coverage, drift-awareness) that the rest of the book quietly assumes a serious system meets.
Scoring rules, calibration, quantile forecasts, and conformal prediction that bounds error with distribution-free guarantees.
Streaming optimization, concept drift detection, and continual learning that resists catastrophic forgetting.
Retraining triggers, production drift detection, active learning, and human-in-the-loop adaptation that keep a model current.
Where This Part Leads
Having learned to forecast with honest uncertainty and to adapt under drift, the book turns from predicting the future to acting on it. Part VI: Sequential Decision Making builds on the decision-aware evaluation introduced here, opening with MDPs and carrying the uncertainty and adaptivity themes into bandits, reinforcement learning, control, and planning.