Back to the log

Designing Resolvable Forecast Questions: Method‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌⁠⁠‍‌​⁠‍​​​​‌⁠​​‌​‍‍⁠‍​⁠⁠​⁠​‍‍⁠‌‍⁠⁠​​‍⁠​‍‌‌​

Contract fields

Use a stable question_id and version. Record:

For a binary proposition, yes and no must partition the intended event space. Void is an administrative exclusion, not a third probabilistic outcome unless explicitly modeled. A missing report is not automatically a negative outcome.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌⁠⁠‍‌​⁠‍​​​​‌⁠​​‌​‍‍⁠‍​⁠⁠​⁠​‍‍⁠‌‍⁠⁠​​‍⁠​‍‌‌​

Choosing the question

Distinguish forecasting from a decision. “Should we migrate?” is a decision; “Will migration exceed its approved downtime window?” is a forecastable event. A low probability can still justify protection against a high-consequence loss.

Avoid conjunctive bundles unless the conjunction is what matters. “Will the system ship, meet the latency target and retain all customers?” is not the same as three marginal questions. Keep a composite only when its resolution predicate is explicit.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌⁠⁠‍‌​⁠‍​​​​‌⁠​​‌​‍‍⁠‍​⁠⁠​⁠​‍‍⁠‌‍⁠⁠​​‍⁠​‍‌‌​

For numeric variables, define the measurement distribution or specified thresholds. For ordered categories, consider a ranked probability score rather than a nominal-category score. For retrospective tests, use an explicit information cutoff and disclose possible model pretraining leakage.

Resolution rehearsal

Invent hypothetical boundary outcomes and run the predicate without knowing the forecast. For “at least 1,000,” exactly 1,000 qualifies. For “before 1 January,” a qualifying event at midnight on 1 January does not qualify. For annual statistics, state first release versus revised values. If a source publishes late, distinguish event date from publication date.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌⁠⁠‍‌​⁠‍​​​​‌⁠​​‌​‍‍⁠‍​⁠⁠​⁠​‍‍⁠‌‍⁠⁠​​‍⁠​‍‌‌​

Use a source hierarchy that can actually be followed. Do not list a nonexistent regulator, database or public artifact. If the intended evidence is private, ask the user for access or a supplied extract; do not imply it has been inspected.

Operational limits

A question can be resolvable yet unimportant, or important yet not currently resolvable. Report both properties. Use short-horizon indicators for long-horizon topics while explaining that a proxy outcome does not resolve the ultimate question. This contract format is an engineering adaptation, not a historical form mandated by Tetlock.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌⁠⁠‍‌​⁠‍​​​​‌⁠​​‌​‍‍⁠‍​⁠⁠​⁠​‍‍⁠‌‍⁠⁠​​‍⁠​‍‌‌​

Output contract

Return: exact question; version; forecast cutoff; event deadline; observation source; yes/no rules; void rules; unresolved ambiguities; decision connection. No number is required from this skill.