forecasting-ch22-causallisted

# Chapter 22: The Causal Forecaster ## Scope and intake Use when an intervention effect needs a defended comparison rather than a purely predictive forecast. Do not use for treating pre-period fit, nonsignificant pretrends or plain OLS as proof of causal identification. Ask only for unresolved material inputs: What treatment, alternative and population define the effect? When did treatment start? Why did it vary? Could controls be affected? What else changed at the same time? ## Input contract and additional evidence A regular series with the treated outcome and at least one control series; name every control column in config `controls`; without that key the tool looks for one column literally named `control`. Synthetic control and placebo-in-space need at least two, better three or more, controls. Declare `intervention` (the first treated timestamp) and write down the `identification` argument before running. Minimal **format illustration**, not sufficient training data: ```csv timestamp,control,control_2,control_3,treated 2010-01-01,99.9,80.7,49.1,104.3 2010-02-01,100.3,81.0,52.4,105.1 2010-03-01,100.1,80.9,47.8,104.6 ``` ## Executable interface Exact CLI columns: `timestamp,control,treated[,controls...]`. Supported method controls: `as_of, controls, event_window, frequency, horizon, identification, intervention, placebos, season, seed`. The tool computes difference-in-differences, a pre-period OLS counterfactual on all declared controls, and, with two or more con