forecasting-ch17-probabilisticlisted

# Chapter 17: Living with Probability ## Scope and intake Use when forecast intervals or quantiles need scoring, calibration and decision interpretation. Do not use for calling scenario endpoints calibrated probability bounds without evidence. Ask only for unresolved material inputs: What nominal coverage or quantiles are claimed? Were forecasts issued before outcomes? What horizons and groups need coverage? Is a separate calibration period available? ## Input contract and additional evidence Either a plain series `timestamp,target`, from which the tool builds and checks its own intervals, or intervals you already issued as `timestamp,actual,lower,median,upper`, which it only scores. The plain series needs enough history for training, a calibration block (`calibration_size`, default 36) and a later test block (`test_size`, default 24). Minimal **format illustration**, not sufficient training data: ```csv timestamp,target 2010-01-01,42.1 2010-02-01,44.8 2010-03-01,47.0 ``` ## Executable interface Exact CLI columns: `timestamp,target | timestamp,actual,lower,median,upper`. Supported method controls: `alpha, as_of, calibration_size, frequency, gamma, horizon, origins, pool, season, seed, test_size, transform`. On a plain series the tool selects a model with the companion engine on history before the calibration block, freezes that specification, refits it at every calibration origin to collect residuals by horizon step, and builds split-conformal intervals at level 1-`