A forecast becomes useful when scenario planning converts its uncertainty into conditional action.

The regional dashboard gives a 70% probability of major violence within one year. The operations director wants a yes-or-no answer on advance transport capacity. The country team wants to wait for clearer evidence. A probability cannot make either choice.

The booking decision has a deadline, sunk costs and consequences for people outside the model. Waiting preserves cash while consuming the commercial option that an early warning was meant to protect.

The number becomes dangerous when the planning team treats it as the central assumption. One expected future then absorbs the remaining 30%. Options are optimised against that average, while the conditions that should change the decision disappear.

Bound the forecast before carrying it into a world. A probability describes a target under stated conditions. Strip away those conditions and apparent precision travels much further than the evidence.

Write four fields beside the number: target, starting state, horizon and evaluation period. “Major violence within one year from an existing minor conflict” is a different task from onset in a peaceful country. It also differs from easing after a ceasefire.

Name the resolution rule so later reviewers know exactly which observed outcome counts. Name the evaluation population because one country case cannot reveal whether a probability was calibrated.

For a multi-state path, keep the transition matrix instead of extracting its largest probability. Where the target changed during development, preserve both definitions and the date of change.

The distinction has measured consequences. One conflict model estimated annual transitions among three ordered states: no conflict, minor conflict and major conflict. Its nine-year evaluation found low-to-high escalation easier than peace-to-minor onset, despite projections extending as far as 40 years.

Record the evidence period too. ViEWS retrospective forecasts for 2016–2018 produced few probabilities above 0.75 or below 0.05. Reasonable average calibration therefore coexisted with weak separation between cases. A score dominated by persistence cannot establish equal skill on sudden onset.

Your scenario input should retain the model version, information cut-off and uncertainty range. The forecast line in the assumptions log then remains bounded. A later update replaces it openly instead of silently validating the original estimate.

Carry the whole range into the scenario work, including a plausible lower and upper reading. An averaged planning assumption would erase the variation that the option test needs to expose.

Separate three sources of uncertainty before choosing responses. A single confidence label hides problems with different operational effects. More research may improve one and leave the other two untouched.

Rare-event evidence is the first limit. A historical atrocity dataset found only 36 genocide or politicide cases worldwide between 1955 and 1997. Broadening the target to mass killing increased the usable sample. It also changed the question being forecast.

Model instability is the second limit. The GCRI revision used 22 variables and data for 175 countries from 1991 onward. Its maintainers removed three variables, added one and compared 14 candidate models for a four-year forecast. Those responsible changes break simple comparisons across versions.

Strategic adaptation is the third limit. Actors conceal information, misrepresent intentions and alter behaviour after a warning. A study of horizon effects also found 59% of strategic-intelligence forecasts were under six months and 96% under one year. Results from those questions cannot settle performance beyond five years.

Put each limit in its own assumption field. Rare evidence argues for option preservation. Model instability argues for sensitivity tests and versioned updates. Strategic adaptation argues for rival pathways, discretion and signposts that detect reactions to your own move.

Add a response beside each field, rather than using confidence as a general discount. The record then shows which uncertainty changes research, contract design, monitoring or decision authority.

A source gap needs collection; a structural break needs a wider range and another world.

Fix the decision before collecting uncertainties. Driver workshops fail when an open trend scan replaces the choice. The room generates interesting material, but nobody knows which differences should change Monday’s commitment.

State the focal decision, system boundary and horizon first. Here the decision is how much transport capacity to secure, for which personnel and routes, through the forecast’s one-year horizon. Existing travel-risk arrangements shape feasible commitments. They cannot decide how much capacity to buy.

Your boundary includes staff, dependants, carriers, airports, land crossings and the approval route for expenditure. Suppliers beyond that boundary remain external assumptions, with named observations where their behaviour could still defeat the choice.

Write exclusions beside the boundary, because hidden scope changes make later comparisons meaningless.

Now collect and rank at least 10 critical uncertainties by importance and uncertainty. Keep predetermined elements separate. Preserve disagreements in the source and assumption log. A 2 × 2 importance-uncertainty map helps sorting, while poor axes can still create implausible combinations.

Include the forecast as one bounded input, alongside political, commercial and operational drivers. Do not let its percentage rank every other uncertainty automatically. The question is which uncertain causal relationships could defeat or enable the live options.

Rank drivers only after you can show participants the evidence, disagreement and dependency behind each one. Two correlated uncertainties should never become independent axes merely because their labels sound different.

Write four causal worlds before testing the favourite. Early option testing bends the stories around the current strategy. The favoured plan receives capable actors and recoverable disruption. Other worlds receive labels and catastrophe.

Use the same frame for all four worlds: starting conditions, branch, actor motives, institutional response, distributional effects, horizon conditions and decision implications. Develop every quadrant to comparable resolution. The UK worksheet illustrates 60 minutes for describing four futures and 30 minutes for naming them; those timings are facilitation examples.

Make the worlds explain divergence. In one, deterrence holds while border processing degrades. In another, violence rises but commercial carriers remain. A third combines intervention with spoiler attacks. A fourth contains rapid escalation and suspended departures. Each path must show why those conditions can coexist.

Challenge the complete set before any option enters. Cedefop uses five tests: plausibility, internal consistency, challenge, engagement and usefulness. Repair a missing actor or correlated axis. Keep an unwelcome development when its causal path survives scrutiny.

Give challengers the original assumption log and ask them to trace each important branch. A world fails comparison when the same carrier is adaptive there but passive everywhere else without reason.

Predetermined elements must appear in every world, even when their consequences differ across pathways.

Wind-tunnel live options through four passes. A workshop vote usually rewards the option whose assumptions sound familiar. Cross-world comparison forces the same objective and decision criteria onto every choice.

First, introduce the four scenarios without ranking them. Second, define each live option and objective. Third, test every option in every world. Fourth, review the implications across worlds. A conventional four-option set creates a 4 × 4 record with 16 option-world cells.

Ask what enables the option, what defeats it, which modification repairs it and which unintended effect follows. Work across all four worlds before changing the criteria. Then record the choice as common, conditional, hedged or rejected.

Common means the same commitment survives every world under the agreed objective and constraints. Conditional means a named observation releases, expands, pauses or reverses the commitment. Hedged means paying to preserve several paths while evidence develops. Rejected means its repeated failure survives reasonable modification, even if the current forecast favours it.

Capture the reason for every cell judgment, since a label alone cannot guide redesign.

A small reservation fee across every world may be common. A larger capacity tranche may become conditional when scheduled carriers stop accepting new bookings. Contracts with several departure points may hedge route closure. A plan requiring continued airport access is rejected when no affordable modification survives the four worlds.

Give every condition an observable signpost and owner. Conditional language without monitoring merely postpones the decision. The same drift returns when nobody owns the observation or the authority to change the option.

Write conditions that independent reviewers can verify. “Scheduled carriers cancel service and stop accepting new bookings” is usable. “Regional tension rises” restates the judgement. Record source, direction, review rhythm, competing interpretation, disconfirming evidence, owner and the decision reopened.

The owner needs access to the source and authority to place the observation before decision-makers. Set a review rhythm that can beat the option deadline, then specify a deputy and escalation route.

An owner should report missed observations as explicitly as fired signposts at each review.

Treat action after warning as part of the causal record. If advance booking keeps a carrier on the route, the forecast and intervention now interact. If diplomatic pressure postpones violence, the observed non-event cannot alone prove forecast error. Score the bounded forecast over its stated evaluation set, and evaluate the intervention through a separate causal review.

Preserve the sequence of forecast, decision, intervention, actor reaction and observed outcome. That timeline lets evaluators test calibration without pretending the organisation remained outside the system it changed.

The final handoff joins six items: forecast record, assumption log, four worlds, option comparison, decisions with owners, and signpost register. You have preserved uncertainty when each option changes for a named observation. The forecast has then done useful work without pretending to choose.

The sponsor should receive the decision record in the ordinary forum that controls spending.