How Over / Under Analyzes Prediction-Market Contracts

What goes into a Deep Analysis, what comes out, and what the confidence score is actually measuring. The methodology page for the platform, written for traders, journalists, and anyone who wants to understand the read before trusting it.

What we’re actually doing

The premise of a prediction market is that the price reflects collective belief about an event. A contract trading at 62¢ implies the market estimates a 62% chance. That premise is robust at the population level (in deep, liquid markets) but breaks down in three common ways:

• The market may be too thin for the price to mean anything, a few traders moving small dollars set the line.
• The resolution criteria may be ambiguous in ways the price doesn’t reflect,“Will X happen by year-end?” can hinge on what counts as X.
• Real-world evidence may have moved since the price was last meaningfully repriced, a poll, a release, a result that the market hasn’t priced in yet.

What Over / Under does, on every contract, is read all of that, the rules, the price, the volume, the relevant evidence, and produce a written read on whether the line is sharp, mispriced, or weakly supported, with a confidence score on the read itself. We don’t predict outcomes. The market is the prediction; we’re a quality check on the price.

What goes in

Three categories of input feed every analysis:

Contract metadata.

The full resolution rulebook from the venue, the exact criteria, the official source named for resolution, the expiry date and time, any tiebreakers, and the secondary rules that govern edge cases. This is the single most important input. A trader who knows the rules cold has a structural edge over a trader who reads the headline; we surface the rules so that edge is available to everyone.

Live market data.

Current bid and ask, recent price movement, traded volume on each side, and order-book depth where available. We treat thin volume as a signal that the price means less, and we explicitly flag low-liquidity contracts in the confidence score.

Outside evidence.

For each contract, we pull in a curated set of public information relevant to the question, the kind of evidence a thoughtful trader would consult before sizing a position. The exact mix varies by category, drawing from authoritative public sources appropriate to the contract type.

The principle is that the model should see what a researcher would consult, not rely on its training-data recollection of the topic.

What comes out: the Deep Analysis

Each Deep Analysis is a structured read on a single contract, with the same four parts every time:

The verdict.

A short, plain-English statement of whether the current price looks well- supported, mispriced (in either direction), or weakly supported because the evidence is thin or the criteria are ambiguous. The verdict is the headline, the part you read first.

The reasoning chain.

Two to four paragraphs walking through what the evidence actually says, what the market price is implying, and where the gap is (or isn’t). This is the heart of the read, it’s also where the analysis is most useful even when you disagree with the verdict, because it surfaces the assumptions and the data the read is built on, so you can override them with your own view.

The resolution source, in plain English.

Every contract resolves on a specific, named source, a NOAA station, a BLS release, a certified state result, an awards ceremony, a price index. We surface this explicitly, in plain English, so it’s clear what you’re actually trading. This is a sneaky-important input: many disputes between traders and venues come down to misreading the resolution criteria, and the rulebook is buried in legal language. We translate it.

The confidence score.

One, two, or three dots. The score is not a probability, it’s how confident we are in the analysis itself, given the strength of the evidence, the clarity of the rules, and the depth of the market. More on this below.

What the confidence score actually measures

The score is built from three components:

• Evidence strength: how much relevant, recent, authoritative information exists, and how consistent it is. A contract on a well-tracked event with multiple converging sources scores higher than a contract on a niche question with one ambiguous source.
• Rule clarity: how cleanly the contract resolves. A contract that resolves on a single named release scores higher than one whose resolution depends on subjective judgments or ambiguous criteria.
• Market signal: how much real activity is in the market. We use a hybrid measure that weighs both recent traded volume and price stability, so a market with both genuine flow and a stable price counts more than one with either alone.

All three need to align for a three-dot read. A market with great evidence but ambiguous rules, or sharp rules but no liquidity, will rate lower, because the analysis itself is structurally less reliable, regardless of what the evidence happens to say.

What we don’t do

A few things worth being explicit about, because they’re what people often assume when they hear “AI prediction-market analysis”:

• We don’t predict outcomes. We analyze whether the existing market price is well-supported. The prediction is the market’s; the read is ours.
• We don’t use insider information. Every input is from public, verifiable sources. There’s no proprietary signal you can’t check yourself.
• We’re not a market maker or sportsbook. We don’t set lines and we’re not the counterparty for trades on Kalshi or Polymarket.
• This isn’t financial advice. Trading prediction markets carries risk. Our analysis is a tool to inform your decision, not a recommendation to act.

Frequently asked

Are you predicting outcomes?

No. We analyze whether the current market price is well-supported by available evidence, not whether the event will happen. A high-confidence read on a 70% market line says the evidence is consistent with that price, not that the event will occur 70% of the time.

What model do you use?

We use a large language model to read the contract context, the live market data, and the curated outside evidence, and to produce the written analysis. We don’t name a specific provider on the page because the stack evolves; the principle is that the model is a reasoning layer, not an oracle of its own, its job is to weigh public evidence, not to invent forecasts.

How often does the analysis refresh?

Live market data updates continuously throughout the trading day. Each contract’s Deep Analysis is regenerated on a regular cadence and also when material new information arrives, so the read you see reflects the current evidence and the current price.

Why don’t I see analysis on every market?

We cover the markets where there’s enough volume, evidence, and clarity for a meaningful read. Brand-new markets, thinly-traded contracts, and markets with extremely ambiguous resolution criteria may not have a published Deep Analysis until they meet that threshold, or may have one with an explicit low-confidence flag.

Is this financial advice?

No. Over / Under produces analysis, not recommendations. Trading prediction markets carries risk, including loss of your principal. Make your own decisions and trade only what you can afford to lose.