A forecast needs a subject, an outcome, a horizon, and a resolution rule agreed in advance. An unresolved question is not automatically a negative outcome.
Start with the decision the client faces
A broad question such as whether the labour situation is a problem does not specify what would count as an answer. Begin with the decision it informs, then name the relevant asset and observable outcome.
A fictional example asks whether Terminal A experiences a continuous work stoppage of at least 48 hours during a stated month. The wording should make clear which activities count, how the duration is established, and what evidence can resolve the question.
Write the evidence rule before the outcome
Specify the resolution date and acceptable evidence at commissioning. Also define how late reporting, conflicting sources, ambiguous duration, or unavailable evidence will be handled. No report of an event does not necessarily mean that the event did not happen.
Freeze the wording once the question is registered. If the scope must change, create an identified revision or a new question and preserve the original. Otherwise the result can become easier to fit after the fact.
Preserve the prior and every revision
The programme proposes starting judgments based on relevant historical evidence, expert input, and analyst reasoning. Record disagreements and the reasons for the chosen prior. An archive of articles does not automatically supply a clean historical event base rate.
For each later update, preserve the forecast timestamp, evidence reference, model or rule version, and explanation. Hand-set weights should be described as judgments to evaluate, rather than presented as learned or validated likelihood ratios.
Evaluate against a stated reference
For a binary outcome, a Brier score uses the squared difference between the forecast probability and the recorded outcome. Lower values indicate lower error under that scoring rule. A skill comparison also needs a reference forecast; a score on its own does not establish useful predictive skill.
Define which forecast timestamps will be scored before evaluation. Treat questions consistently so that issuing more updates does not give one question disproportionate influence. Keep unresolved cases visible, state the evaluation sample, and retain unsuccessful forecasts. The Venezuela programme does not yet claim a validated track record.
What to watch
- Is the outcome precise enough for an independent reviewer to resolve?
- What happens if the named evidence is unavailable or contradictory?
- Which forecast timestamps and reference forecast will be used for evaluation?
Sources & editorial notes
A programme-design guide adapted from the September 2026 working plan. Examples describe intended methods, not current incidents, confirmed partnerships, or operating field coverage.
Prepared 11 September 2026. Read our editorial standards. To suggest a correction, contact the editorial desk.