How to Cite Prediction Market Data in Research (Without Getting the Stats Wrong)

TL;DR

• A prediction market price is a timestamped point estimate: cite "Kalshi priced the contract at 63% as of March 3, 2026, 14:00 UTC," never "there was a 63% chance."
• A single trade price carries no built-in margin of error. Build one from order-book spread and depth, or report the market's liquidity alongside the price.
• Resolved-market datasets overstate calibration through survivorship bias: markets that got delisted, disputed, or lost liquidity before resolving drop out of the sample, and they don't drop out randomly.
• A live price is a forecast; a resolved outcome, paid at $1 or $0, is the only ground truth a calibration study can use.

Cited a Polymarket price in a paper and had a reviewer ask why there's no confidence interval attached? That's usually the moment a researcher realizes a market price isn't the same kind of number as a poll result.

A prediction market price is a snapshot of aggregate belief at one moment, not a measured probability with a confidence interval. Cite it as "the market implied a 34% probability as of March 3," not as "there was a 34% chance." That distinction is the difference between a defensible citation and a methodology error a peer reviewer will catch.

Researchers increasingly pull odds from Polymarket, Kalshi, Limitless, Predict.Fun, and Opinion the way they used to pull polling averages. The instinct is right. The execution usually isn't, because market prices behave differently from survey data in ways most academic training never covers.

What Does a Snapshot Price Actually Represent?

A price on a binary contract reflects what traders were willing to pay at that instant, given the information and liquidity available to them right then. Say a contract trades at $0.63. That's a 63% implied probability, but it's the market's estimate conditioned on everything known up to that second, not a stable population parameter you can treat like a poll's margin of error.

Two things follow from this, and both matter for citation practice.

First, the price moves. A contract at 63% on Monday can sit at 40% by Thursday without any of the past estimates being "wrong." Each was correct given its information set at the time. If you're citing a prediction market price in a paper, timestamp it. "Kalshi priced the contract at 63% (March 3, 2026, 14:00 UTC)" is a citation. "Kalshi shows 63%" is not, because by the time a reader checks, it won't.

Second, a single price is a point estimate with no reported variance. Polls come with sample size and margin of error baked into the methodology section. A market price comes with neither, unless you compute it yourself from the order book. If your paper needs a confidence interval, you build it from spread and depth data, not from the last trade price alone.

Why Should You Treat Liquidity Like a Sample-Size Problem?

Treat thin liquidity the way you'd treat a small sample: as a caveat that limits what you can claim, not as a defect to paper over. A market with a few hundred dollars in open contracts can be moved by one trader with a strong opinion and no informational edge. A market with deep, sustained volume across many participants is aggregating more independent judgment.

Neither Polymarket, Kalshi, Limitless, Predict.Fun, nor Opinion publishes a standardized liquidity score you can drop into a regression, so you have to build your own threshold. Reasonable approach: pull the market's total volume and the depth at the current price (how much can be traded before price moves materially), and set a minimum below which you exclude the market from your dataset or flag it separately. Document that threshold in your methodology section. A reviewer will ask why a $400 market and a $4 million market are weighted the same if you don't.

This is also where cross-platform comparison earns its keep. If Polymarket and Kalshi are pricing the same event five points apart, that gap is often explained by one platform having ten times the depth of the other. The divergence is data, not noise, but only if you report the liquidity alongside the price.

Should You Cite a Live Price the Same Way as a Resolved Outcome?

These are two different kinds of evidence, and conflating them is the most common mistake in market-derived research.

A live price is a forecast. It tells you what informed capital believed at a point before the outcome was known. It's useful for studying belief formation, sentiment shifts, or how new information gets incorporated. It is not evidence about what actually happened, and it should never be cited as if the event's probability was later "confirmed" by the price alone.

A resolved outcome is a fact. The contract paid $1 or it paid $0. That's ground truth for calibration studies, where you're checking whether markets that priced events at 70% actually resolved yes about 70% of the time across a large sample.

The error researchers make is pulling only resolved markets to build a calibration dataset. That sample is not representative of markets in general, and here's why.

How Does Survivorship Bias Distort Resolved-Market Datasets?

If you only look at markets that resolved, you're only looking at markets that ran their full course, on platforms that stayed operational, in categories liquid enough to attract closing volume. Markets that got pulled for ambiguous resolution criteria, categories a platform quietly stopped supporting, or events that lost all trading interest before resolution tend to drop out of "resolved markets" datasets entirely, and they don't drop out randomly.

This mirrors the survivorship bias problem Brown, Goetzmann, Ibbotson, and Ross documented in mutual fund research: tracking only funds that still exist overstates average performance, because the funds that got shut down or merged away for underperforming already left the sample.

Apply the same skepticism to resolved markets. A dataset built only from resolved contracts skews toward events that were well-specified, liquid, and uncontroversial enough to resolve cleanly. Markets with vague settlement language or thin follow-through, arguably the ones most informative about market limitations, are underrepresented.

The fix is not complicated, just tedious: pull the full universe of markets you intended to study, including the ones that were delisted, disputed, or expired without resolving, and report that attrition rate. "Of 340 markets meeting our criteria, 61 did not resolve or were excluded for disputed settlement" is a sentence that makes your calibration claim credible. Omitting it is a sentence a careful reviewer will notice by its absence.

How Do You Build a Defensible Research Dataset?

• Record the timestamp and source platform for every price you cite. A price without a timestamp is not a citation, it's an anecdote.
• Capture liquidity alongside price: volume and depth at the time of the snapshot, not just at market close.
• Separate live prices from resolved outcomes in your data structure. Never let a probability estimate and a binary outcome sit in the same column without a flag distinguishing them.
• Track markets that didn't resolve, not just the ones that did. Attrition is data.
• Cross-check across platforms where the same event trades on more than one. Reporters run into the same verification problem before a number goes in a story; our journalism workflow piece covers that side of the same discipline. Comparing Polymarket, Kalshi, Limitless, Predict.Fun, and Opinion on an overlapping event gives you a rough sense of estimate stability that a single platform can't.

Sources

• Wolfers, Justin, and Eric Zitzewitz. "Interpreting Prediction Market Prices as Probabilities." NBER Working Paper No. 12200. nber.org
• Brown, Stephen J., William N. Goetzmann, Roger G. Ibbotson, and Stephen A. Ross. "Survivorship Bias in Performance Studies." terpconnect.umd.edu
• Kalshi Help Center. "Kalshi API." help.kalshi.com
• Polymarket Documentation. "Data API v2 Overview." docs.polymarket.com

PredictionHero aggregates publicly available prediction market data for informational purposes only. This is not financial advice. Prediction markets may not be available in all jurisdictions.