How to write prediction market questions and resolution criteria

Write the question and its evidence test together. A useful proposal says exactly what counts, which source will establish it, and which dates matter. Then try to break it with borderline cases before you submit it.

The goal is not a longer rulebook. It is a question that two readers can apply to the same evidence without silently making different assumptions. A precise question can still be a poor choice if its metric does not measure what you actually want to know.

This guide takes one fictional idea from vague wording to a filled proposal and test sheet. It is a drafting exercise, not HypaForecast's binding market rules or a ready-to-publish market. For the basic mechanism, see prediction markets explained.

Start with the decision, not the headline

Ask what someone would learn from the answer. “Will the festival succeed?” might mean a large audience, satisfied visitors, a balanced budget or a repeat event next year. Those are different questions.

Choose one observable outcome and explain its limits. An admissions total can measure entries, but it cannot establish visitor satisfaction or profitability. If your real interest is satisfaction, adding more detail to an admissions question will not fix that mismatch.

Metaculus's question-writing guide gives substantial advice on definitions, evidence and edge cases. Cultivate Labs' guide also includes an annotated worked example. The exercise below applies those general principles to a new proposal, then tests whether its own clauses actually answer difficult cases. Neither source's platform procedures become HypaForecast rules.

Before and after: make the measurable claim visible

Fictional exercise: Riverlight Science Week, its organizer, report, dates and all results below are invented. There is no real event or source URL behind this example.

Version	Proposed wording	What changes?
Before	Will Riverlight Science Week be a big success this spring?	“Big success” has no single test; “this spring” gives no exact window or timezone.
Better title	Will Riverlight Science Week record at least 10,000 admissions during May 1–7, 2030 (UTC)?	Names a metric, inclusive threshold and dated observation period.
Necessary companion	Use the qualifying organizer report and revision cutoff defined below. Cancellation and missing evidence remain open design questions.	Makes clear that the title alone is not the complete proposal.

“At least 10,000” includes exactly 10,000. “More than 10,000” would not. Do not switch between those phrases in the title, description and criteria.

The improved title also narrows the idea: it asks about recorded admissions, not whether the festival deserves to be called successful. Keep that distinction in the background explanation so readers do not mistake the measurable proxy for the larger claim.

A filled prediction market proposal worksheet

Here is a draft you can inspect rather than a blank checklist. The clause labels make it possible to trace each test back to the proposed rule.

Status: NOT READY for publication. The ordinary numerical cases are specified; cancellation, missing evidence and other listed exceptions still need a valid treatment agreed through the applicable platform review process. These are deliberately exposed gaps, not instructions to resolve a live market.

Clause	Field	Filled proposal — fictional throughout
A	Question and scope	Will Riverlight Science Week record at least 10,000 admissions during May 1–7, 2030 (UTC)? Only the named 2030 edition counts.
B	Observation period	Include admissions from May 1, 2030 at 00:00:00 UTC up to, but not including, May 8, 2030 at 00:00:00 UTC.
C	Metric and unit	Use the report's whole-number `total_admissions` field for the complete period in B. Its proposed definition is counted entry scans, including repeat visits; exclude staff scans, online views and ticket reservations without entry. Unit: admissions, not unique people. Do not round or estimate a missing value.
D	Primary evidence	The fictional Riverlight Science Week organizing office's “2030 Attendance Report,” identified by edition and observation period, with status `final`. A real proposal must provide the actual public URL, field definition and accessible revision history; none is invented here. News stories and promotional posts are not substitutes in this exercise.
E	Version and cutoff	Of qualifying final revisions publicly issued on or before May 10, 2030 at 12:00:00 UTC, use the one with the latest publication timestamp. Preserve its revision identifier and timestamp. A later qualifying revision before the cutoff replaces an earlier one. Corrections published after the cutoff do not replace the selected revision in this exercise.
F	Ordinary YES test	If a qualifying report has been selected under D–E and no exception in H applies, YES when its complete-period `total_admissions` is 10,000 or greater.
G	Ordinary NO test	Under the same conditions, NO when that field is a nonnegative integer below 10,000. Absence of a qualifying report is not a value of zero.
H	Unresolved exceptions	No qualifying report; an inaccessible or withdrawn selected report; conflicting final revisions with the same latest timestamp; a missing, non-integer or negative field; changed counting definitions; cancellation, postponement or a shortened event. These require clarification before publication. This worksheet assigns no outcome, refund or void treatment to them.
I	Proposed forecast/trading close	April 30, 2030 at 18:00:00 UTC, subject to the actual platform's review. This illustrative close is separate from both the observation period and the evidence cutoff.
J	Determination and settlement	Do not treat the close or the evidence cutoff as a promised settlement time. The actual platform must specify its applicable determination, review and settlement process; this worksheet supplies no substitute authority or timing guarantee.

The source field in C is doing important work. A ticket reservation is not an entry, and one person visiting twice is not two unique visitors. Before borrowing this structure for a real topic, confirm that your chosen source actually reports the metric you have defined. A detailed description of a field that does not exist is not a usable evidence plan.

Clause E makes a deliberate choice: the result follows a specified reporting version, not all facts that might become known forever. That choice needs to fit the intended question and be accepted for the actual market. If you instead want the first release or a later corrected figure, write that explicitly and rerun the tests. Do not change a live market's criteria yourself.

Keep the clocks separate

A market can stop accepting forecasts or trades before the underlying event is over. The evidence can arrive later still. Write each time separately rather than using “deadline” for all of them.

Proposed close
April 30, 2030 · 18:00 UTC
Forecast/trading cutoff proposed for review.
Observation window
May 1 · 00:00 UTC → May 8 · 00:00 UTC, 2030
Start included; end excluded. Only admissions in this window count.
Evidence cutoff
May 10, 2030 · 12:00 UTC, included
Select the latest qualifying final revision issued by this time.
Determination and settlement
Timing not specified by this exercise.
Follow the actual platform's applicable process.

Fictional proposal timeline, not a live market schedule. The gap between observation and reporting allows evidence to arrive; it is not an extension of the event window.

Suppose an admission occurs on May 7 at 23:59:59 UTC and appears in a qualifying report published on May 9. It is inside the observation window even though the report came later. An admission on May 8 at 00:00:00 UTC is outside the window, however quickly it is reported.

That distinction changes the question you are asking. “Did it happen by Friday?” and “Was it publicly reported by Friday?” can have different answers. Decide which one you mean before selecting a source cutoff. For reading an existing market rather than drafting one, use the rules-and-timing steps in how to trade prediction markets.

Try to break the proposal: an adversarial test sheet

For the ordinary numerical rows below, assume the event ran as specified, the selected report is accessible, and no H exception applies. All timestamps are UTC and all scenarios are fictional. “YES” and “NO” mean the answer under this exercise's draft clauses—not a platform settlement instruction.

Test	Fictional evidence	Result under this draft	Clause and reason
Threshold equality	Selected report records 10,000 admissions.	YES	F includes equality.
Just below	Selected report records 9,999 admissions.	NO	G covers nonnegative integers below 10,000.
Different unit	A press release says “10,500 tickets reserved”; the qualifying report records 9,800 admissions.	NO	C–D use admissions in the report, not reservations in a press release.
Event before boundary, report afterward	An admission at May 7, 23:59:59 appears in a qualifying May 9 report.	Admission is within B; total still decides YES or NO.	B limits the event time, while E limits publication time. One qualifying admission does not establish the threshold.
Event exactly at excluded boundary	An entry scan occurs at May 8, 00:00:00.	Does not count toward the period.	B excludes that instant. A report that folds it into the period total would need clarification rather than an improvised subtraction.
Correction exactly at cutoff	Revision 1 says 10,020; a qualifying final revision 2 published May 10 at 12:00:00 says 9,980.	NO	E includes the cutoff instant and selects revision 2; G then applies.
Correction after cutoff	Revision 1 says 10,020; the correction to 9,980 is published May 10 at 12:00:01.	YES under the illustrative version rule.	E retains the qualifying pre-cutoff revision; F applies. This assumes the selected report has not been withdrawn and no H exception applies.
Only report arrives late	The first final report, showing 10,020, appears May 10 at 12:00:01.	Unresolved design; not NO.	D–E provide no qualifying report, so H applies.
Source disappears	The selected report cannot be accessed.	Unresolved design.	H requires a verified fallback or other applicable treatment before publication; the author cannot invent one after the fact.
Event cancelled or moved	The organizer cancels the event, or moves it to June.	Unresolved design, even if a notice mentions zero May admissions.	H explicitly takes precedence over the ordinary numerical tests. No automatic NO or refund is assigned.
Competing revisions	Two final revisions share the latest timestamp but give different totals.	Unresolved design.	H leaves no tie-break rule. Selecting the preferred total would add a new rule.

The late-report pair is especially useful. A correction arriving after the cutoff does not replace an already qualifying revision under E. But a first report arriving after the cutoff leaves no qualifying revision at all. The same late timestamp does not produce the same result in those two situations.

Do not hide an unresolved row by writing “otherwise NO.” That phrase would silently turn a missing report into evidence of failure. Instead, list the gap, investigate a dependable fallback, and ask the platform to review the proposed treatment. If no reliable evidence plan is available, choose a narrower question or a different topic.

Turn your worksheet into a reviewable proposal

Give a second reader the title, clauses and a few test cases without your intended answers. Ask them to identify the source version, counted period, applicable clause and result. If their answer differs, find the missing definition rather than arguing about what the headline “obviously” meant.

Before submitting, check three things:

Meaning: the title and metric answer the same question. Entries are not unique visitors; a reported total is not a claim about unreported reality.
Evidence: the real source exists, covers the entire period, uses the stated unit and has a workable version history. Any fallback has its own selection rule, not just “another reliable source.”
Exceptions: every plausible test has either a proposed treatment consistent with the platform's rules or an openly identified issue for review. A worksheet with unresolved exceptions is not ready to publish.

For HypaForecast, read the current Creator program information for the appropriate next step. HypaForecast reviews proposed markets before publication, including their questions, outcomes and resolution criteria. Preparing this worksheet does not guarantee acceptance or give you authority to create a live market or decide its result.

Frequently asked questions

Can the title contain all the resolution criteria?

Keep the title readable, but do not let it promise something different from the full criteria. Put the main event, threshold and period in the title where practical; use the accompanying rules for source selection, revisions and exceptions. If a short title would mislead readers, rewrite the question rather than burying the difference.

Should missing evidence mean NO?

Not automatically. In this example, NO needs a qualifying report with a value below 10,000. No report means that test cannot be applied. A real market needs an applicable rule for missing evidence; this worksheet does not supply a default outcome, refund or void policy.

Should later corrections change the answer?

It depends on the rule chosen before publication. This exercise selects the latest qualifying final revision issued by a fixed cutoff and excludes later corrections. That is one hypothetical design, not a universal policy. Test whether it answers the question you intend, and confirm the actual market's rule before relying on it.

Does precise wording eliminate disagreements?

No. It can make disagreements easier to locate, but unexpected events, changing definitions and unavailable evidence can still cause problems. The useful deliverable is a clear proposal with tested boundaries—not a promise of dispute-free resolution.

Explore HypaForecast markets

Browse the market list, or inspect the question and current round details for the 5-minute and 15-minute BTC markets.

Sources and scope

Published by HypaForecast. Source information checked October 26, 2026.

Metaculus — Question writing and submission guidelines: general drafting guidance on definitions, evidence, edge cases and dates. Metaculus-specific procedures are not applied to HypaForecast here.
Cultivate Labs — The Anatomy of a Forecast Question: question components and an annotated worked example.
HypaForecast — Creator program information: current program and proposal-review information.

The Riverlight worksheet, timeline and test results are original hypothetical teaching material. They do not describe an actual market, measured attendance or an approved settlement policy.