forecasting-ch11-crowdslisted

# Chapter 11: The Crowd and the Ox ## Scope and intake Use when several independent quantity estimates need aggregation with attention to common bias and outliers. Do not use for assuming more respondents or more AI-generated opinions automatically create more independent evidence. Ask only for unresolved material inputs: Do estimates share a target, horizon and units? Who saw whose answer? Which sources are common? Are there past resolved questions for testing weights? ## Input contract and additional evidence CSV question,expert,estimate,actual; one independent initial estimate per expert/question, actual finite for CLI scoring; keep unresolved estimates separately. Document units and any elicitation order. Do not mix probability and quantity estimates. Minimal **format illustration**, not sufficient training data: ```csv question,expert,estimate,actual Q1,E1,90,100 Q1,E2,110,100 Q1,E3,1000,100 ``` ## Executable interface Exact CLI columns: `question,expert,estimate,actual`. Supported method controls: `as_of, extremize_a, frequency, horizon, season, seed`. The adapter requires resolved question actuals and compares fixed mean, median and 20%-each-tail trimmed mean. When the estimates are event probabilities with binary actuals, `extremize_a` (a positive number, 2.5 is the usual starting value) adds a logit-extremized pool and scores every rule by Brier as well as MAE; extremization helps only when experts share information and the plain pool is too timid, and the