How we research a market

What an analysis is, and what it is not

An analysis is a structured basis for your own decision. It is not a buy or sell recommendation, and it is not investment advice. Every report states a probability we derived ourselves and the price the market was showing at the time, so you can see where the two disagree and judge the reasoning behind the gap.

The report is generated. A person does not review each one before it appears. That is the single most important thing to know when reading it, and it is why every factual claim carries its sources and why the sections below describe exactly what the model is asked to do.

The four research steps

The model is instructed to work through the same sequence on every market, rather than summarising whatever a search returns first:

• Identify the two to four empirical questions that actually decide the outcome.
• Research each one against recent reporting and authoritative primary sources — official agencies, filings, polls, schedules, and the resolution source the contract itself names. Developments from the last seven days are searched for explicitly.
• Establish a base rate: how often have events of this type resolved YES historically?
• Weigh conflicting evidence rather than picking a side, and carry the remaining uncertainty into the confidence rating and the probability.

How the probability is calibrated

The market price is the starting point. A liquid market aggregates the views of people with money at stake, and measured against resolved outcomes it has been hard to beat — so the estimate moves away from it only for specific, sourced evidence the market has plausibly not priced in yet, and the report names that evidence and states how far it moved and why. Without such evidence, the estimate stays within about three points of the price, and the report says it found no reason to deviate.

Two checks sit on top. The model is told to guard against a lean towards YES, because questions of the form "will this happen by a date" resolve NO more often than a compelling story suggests. And where our own corpus of resolved markets supports it, the report is given how often markets in the same topic and price range actually resolved YES — as a check on whether prices in that range have run high or low, never as the answer for the market at hand.

Time remaining is part of the calculation: the closer a deadline comes without the required event having happened or been scheduled, the further the estimate has to move toward the base rate for last-minute resolution. Round numbers are discouraged — 62 and 87 are acceptable answers, a reflexive 50 or 80 is not.

The resolution rules get read like fine print

A large share of surprises in prediction markets come from the rule set, not from the world. Every report therefore carries a resolution section that states in one sentence what must happen for YES, which authority decides and when, the exact deadline including timezone where the rules name one, and an ambiguity rating.

Two fields exist purely to catch the common traps: hidden clauses — vague or counterintuitive wording, each with a short note on why it could flip the expected outcome — and an explicit comparison between the contract title and the actual rules. When they agree, the field says so rather than staying silent.

Sources and citations

Key facts carry indices into the search results the model actually used, and those citations are rendered visibly under every analysis. The instruction is to use an empty list rather than guess when it is unclear which result supports a claim.

This is the part to read critically. Citation indices are produced by the same model that writes the text; a wrong index is a plausible failure, and the safeguard against it is that the sources are shown to you rather than summarised away. Treat a claim whose source does not support it as what it is — an error worth reporting.

Qualified wallets, not big wallets

Where a report or panel refers to smart money, it means a cohort selected for demonstrated edge, not for size. A wallet has to bring a minimum history of resolved positions before it is scored at all — below that, a handful of lucky calls would produce a perfect record.

The qualifying threshold is measured rather than set: it is derived from the distribution of scored wallets and re-derived as that population changes, so the cohort stays the top slice instead of drifting with a number someone picked once. A verdict also accounts for how many distinct wallets stand behind a position and how recent they are, and it says so instead of inventing a signal when the sample is too thin or one position dominates it.

Market context

For markets with a tradeable underlying, reports are enriched with context computed from public data rather than described in prose: spot price, the move required to reach the threshold, a historical base rate, realised volatility, funding and open interest, and — where the instrument exists — an options-implied probability. Sports markets get schedule and standing context instead.

Macro is the acknowledged gap. Under the constraint of using only sources that need no API key, no provider covers questions like a September rate decision, and the reports say nothing rather than something unsourced.

What this method cannot do

It cannot see private information, and it cannot outrun a market that has already priced in news. It depends on search results, so a market whose decisive facts are not published anywhere is a market where the estimate is weak — the confidence rating is where that shows up.

It is also not a track record. How well the estimates have performed is a separate, measured question; how it is scored is described on the signal scoring page.