Field guide

How forecasting works

How Matakite works in practice: the four question types and their parameters, what a well-formed question looks like, and the technical details behind records, resolution, and scoring. For the why, see About.

Anatomy of a question

Every question, whatever its type, carries the same core fields:

Text: The question itself.
Question type: Binary, multiple choice, numeric, or date, with the type-specific parameters described below.
Resolution criteria: The precise rule that decides the outcome, including the source used to verify it. A reader at resolution time must be able to settle it without judgement calls.
Closes / resolves: When forecasting stops, and when the outcome is expected to be known.
Resolver: Who settles it (the author, by default). Resolution is a public record, not an authority — see below.
Background & fine print: Optional context for forecasters; not part of the resolution rule.
Domain & tags: Optional, and independent of each other. The domain is one curated subject area (politics, sports, tech…) — it’s the axis your track record is broken down by. Tags are free-form labels for discovery. Neither affects the resolution rule.

Writing a good question

A good question resolves cleanly — a stranger reading it later reaches the same verdict you would. Three things do most of the work:

A firm deadline. Say “by 2027-01-01 UTC”, not “soon”.
A named source. Point at who or what confirms the outcome — “per the BLS release”, “per FIFA’s official record”.
No ambiguity. Spell out edge cases in the criteria. If reality can’t be determined under the rule, it resolves ambiguous and no one is scored.

Question types

Binary — yes / no

A single claim that resolves yes or no. You forecast a probability from 0 to 100%.

Parameters — none beyond the core fields.

Examples

Will a crewed vehicle reach the Martian surface before 2032-01-01 UTC, confirmed by a national space agency or independent reporting? Firm date, clear confirmation source.
Will the US CPI year-over-year for December 2026 be at or below 3.0%, per the BLS release in January 2027? Named source, exact threshold and period.
Will the home team win the 2026–27 league opener, per the official final result? Unambiguous single event.

Multiple choice

A fixed set of mutually exclusive options — exactly one resolves. You forecast a probability across the options that sums to 100%.

Parameters — options (2–50 labels), each with a stable id inside the pinned question version. Editing an option label creates a new question version; existing forecasts remain attached to the version they answered. Make the options exhaustive: add an “Other / none of the above” when they might not be.

Examples

Which party will win the most seats in the next general election? — Progressive · Unity Coalition · Green Alliance · Other Add “Other” so the set is exhaustive.
Which film will win Best Picture at the 2027 Academy Awards? — one option per nominee The nominees are the whole set — exactly one wins, so no catch-all is needed here.
Which company ships a sub-$600 consumer AR headset first? — Apple · Meta · Samsung · Someone else · None before 2028 A “none by DATE” option handles the no-winner case.

Numeric — a number in a range

Resolves to a single number within a declared range. You forecast a median estimate, and optionally a p10–p90 band to say how sure you are.

Parameters — min / max (the range); optional unit; logScale for quantities spanning orders of magnitude; and open bounds (openLower / openUpper) when the outcome could fall beyond an end.

Examples

How many named Atlantic hurricanes will form in the 2027 season, per NOAA? — range 0–30, unit “hurricanes” Bounded count with an official source.
What will the Nasdaq-100 closing level be on 2026-12-31, per the official Nasdaq close? — range 5,000–40,000, log scale, open above Log scale suits a wide multiplicative range.
How many total goals will be scored across the 2026 World Cup, per FIFA match records? — range 0–300, unit “goals” Clear counting rule and window.

Date — when will it happen

Resolves to the datetime an event occurs, within a declared window. You forecast a median date, optionally with a p10–p90 band. It’s the numeric type on a calendar axis.

Parameters — min / max (earliest / latest), and open bounds when the event could occur outside the window. A date forecast still requires an event date; if “not by the cutoff” is a possible outcome, use a binary question or an explicit multiple-choice option.

Examples

When will the next magnitude 6.0 or greater earthquake be recorded worldwide, per the USGS Earthquake Catalog? — window from question creation to +30 days, open after The event recurs; the open upper bound covers a wait longer than the initial window.
When will the Federal Reserve next cut the target rate, per an FOMC statement? — window from today to +2 years Bounded, official trigger.
When will the new rail line open to passengers, per the operator’s announcement? — window Q1 2027 to Q4 2029 A named confirming event.

Forecasting & resolution

Revise freely. You can update a forecast as often as you like before close. The Matakite app creates a separate record for each revision, and the appview retains the revision and retraction events it observes. Scoring rewards being right earlier because it is time-weighted.
The crowd stays hidden at first. It is never shown with fewer than 3 forecasters. With at least 3, it reveals once either 5 distinct people have forecast or 24 hours have passed since the question opened, reducing early anchoring.
Resolution records an outcome. Once the question’s expected resolution time arrives, the resolver can record it. Non-binary types also record the winning option, the resolved value, or the resolved date. ambiguous (undeterminable) and annulled (defective question) score no one.
Disputes are public. A resolution can be disputed, and the dispute is itself a record.

Scoring & reputation

Every scoreable question contributes one normalized loss in [0,1] (lower is better), so the four types combine into one headline score:

Binary — Brier score, (p − outcome)², with the displayed percentage converted to a probability from 0 to 1.
Multiple choice — normalized multiclass Brier over the options (a two-option question matches the binary formula exactly).
Numeric / date — a quantile CRPS: how far the whole forecast distribution sat from the resolved value, rewarding honest uncertainty.

Each loss is time-weighted over your forecast trail and scaled by coverage (how much of the open period you were on the record). Your profile also shows calibration diagnostics — reliability curves for binary and multiple-choice, PIT histograms for numeric and date — plus a per-type and per-domain breakdown. The separate leave-one-out crowd comparison provides context for question difficulty; it does not alter the domain’s raw mean loss. Scores are computed under a versioned, public spec (currently v0.3); anyone can recompute them from the records.

A domain where you’re consistently ahead of the crowd earns an “Expert” badge for that area (e.g. “Geopolitics & Security Expert”). The bar is deliberately high — it takes a solid sample and a clear margin (the edge is damped for small samples) — so the badge means the record backs it, not a lucky run.

Moderation & flags

Anyone can flag a question with a reason. Flags are public records; matakite aggregates them per question and, past a threshold, changes how the question is displayed — it never deletes the record:

Annotate — unresolvable or duplicate: shown with a notice.
Collapse — spam: tucked behind a click.
Hide — harassment or privateIndividual: dropped from feeds, with the question body withheld by the Matakite appview even on direct links.

Flags are weighted by the flagger's scored history — the reference policy uses scored-question count and mean normalized loss, with bounded weights. It also requires both a minimum number of distinct flaggers and a minimum total weight; this raises the cost of coordinated bad-faith flagging but cannot make it impossible. Flags are public, and operators can override the resulting display policy. Matakite does not currently provide a separate author appeal action. (A catch-all other flag needs a note to count at all.)

Disputes

A resolution is a public claim, not the last word. If you think a resolver got it wrong, you file a dispute — a public counter-attestation naming the semantic outcome you assert is correct, with optional evidence. It is published in your repository and attributable to your DID. Matakite does not currently adjudicate disputes or apply an automatic penalty to the disputer.

The current dispute record expresses yes, no, ambiguous, or annulled; it does not encode an alternative option, numeric value, or date. Disputes feed the resolver's track record — “N resolutions, M disputed” — which appviews surface so readers can judge how much to trust a given resolver. None of this is an authority; it's public counter-attestation, and everyone can see who claimed what.

Records & the protocol

Records originate in your PDS. Questions, forecasts, resolutions, flags, and disputes are public records in your AT Protocol repository under the app.matakite.* namespace. Repository commits are cryptographically signed; record createdAt values are author-declared metadata, not trusted receipt timestamps.
Forecasts pin the question version. A forecast references the exact version it answered (by content hash), so edits to a question after the fact are detectable and don’t silently rescore you.
The index is derived. Matakite maintains a serving index for questions, aggregates, scores, and observed deletion receipts. It is derived from public records and observed lifecycle events rather than being where users author their records.
Resolution is data, not decree. Any account can publish a resolution record; clients choose which resolvers to trust (the default is the question’s designated resolver).

Verification & your data

Matakite is built on AT Protocol, so your questions, forecasts, resolutions, flags, and disputes live in your own repository, not inside a Matakite account. You can edit or delete them whenever you like. That freedom is a good thing, but forecasting only works if nobody can quietly rewrite yesterday’s call after seeing today’s result. So verified scoring keeps a public history of what Matakite actually saw and when.

That is the auditor’s job. It is not rummaging through your account: it only follows the public Matakite records you choose to publish. Think of it as a scorekeeper keeping the match sheet so everyone can check the result.

A trail anyone can check. The auditor adds receipts to its own public repository. For a new or updated record, the receipt includes a snapshot of the public record as it appeared then. Matakite’s auditor only adds receipts and watches for unexpected changes, and anyone can replay the trail to check the scores.
When it was seen, not your clock. For verified scoring, the auditor’s first-sight time counts, not the createdAt value in the record. That makes changing your clock or backdating a record useless.
Deleting is a withdrawal, not a time machine. A deleted forecast stops being live from that point, while anything already witnessed stays part of the historical record. Questions and resolutions may also need to remain visible so existing forecasts still make sense. The auditor’s public snapshot stays in the trail, so anything you publish should be treated as public and on the record.