forecasting-ch24-structural-breakslisted

# Chapter 24: The Anomaly ## Scope and intake Use when forecast residuals may signal a structural break and adaptation policies need explicit evaluation. Do not use for tuning an alarm threshold on the very break used to claim detection success. Ask only for unresolved material inputs: What stable period can calibrate the detector? What false-alarm burden is tolerable? Which errors are available online? What action follows the first alarm? Could the anomaly be bad data? ## Input contract and additional evidence A regular series with at least 80 observations. `calibration_size` names the opening window believed to be a stable regime; the CUSUM mean and standard deviation come from it, so choose it before looking at the later data. Seasonal series should be deseasonalised first, or the seasonal swings will be dated as breaks. Minimal **format illustration**, not sufficient training data: ```csv timestamp,target 1871-01-01,1120 1872-01-01,1160 1873-01-01,963 ``` ## Executable interface Exact CLI columns: `timestamp,target`. Supported method controls: `as_of, calibration_size, drift_window, frequency, horizon, max_breaks, min_segment, rolling_window, season, seed`. The tool dates mean shifts by binary segmentation with a Gaussian cost and a penalty of three times log(n) times a robust noise variance (about a one percent false split rate on white noise), reports the strongest single split with its Welch t statistic, runs a two-sided CUSUM alarm calibrated by simulation t