The most common question new readers ask about prediction markets is: are they actually accurate? The honest answer is yes, on the right kind of question, and with caveats. Decades of academic and platform-level data show that prediction markets are well-calibrated forecasters on binary, source-resolvable, well-trafficked events. They are not magic, and they fail in predictable ways on the wrong kind of question. This piece walks through the empirical record.
The long-term academic record
Prediction-market accuracy has been studied since the Iowa Electronic Markets started trading in 1988. The broad finding from that literature: on US presidential elections from 1988 onward, the IEM closing-price probability beat individual polls in 75% of head-to-head comparisons within the final 100 days of the race. The IEM was rarely the most accurate forecast available, but it was competitive with the best aggregators (Nate Silver’s models, RealClearPolling averages) and meaningfully better than individual polls.
A 2008 paper by Wolfers and Zitzewitz consolidated this into the canonical academic statement: prediction-market prices are well-calibrated — when the market says 70%, the event happens roughly 70% of the time across a large enough sample. That calibration property is the technical reason markets work as forecasts.
Where Polymarket and Kalshi sit today
Both major modern prediction markets cite their own accuracy data. Polymarket reports 94%+ accuracy on call outcomes in the final month before resolution, across all market types. Kalshi reports similarly high accuracy on its CFTC-regulated event contracts. Both numbers should be read carefully:
- “94% accuracy” usually means the favorite (above 50%) wins 94% of the time when the market closes more than a month before resolution.
- That is a Brier-score-friendly metric, not a forecasting-skill metric. A market that says “favorite wins” is right ~94% of the time on lopsided markets — which is most markets.
- The harder test is calibration on the 50/50 tail: when the market says 60%, does the event happen 60% of the time?
On the harder calibration test, both platforms have generally been well-calibrated. The 2024 US presidential election market closed at roughly 60% Trump on Polymarket and 58% on Kalshi; Trump won. The 2022 midterm Senate market gave roughly 67% Democratic; Democrats held the chamber. The 2024 NFL playoff markets called the Super Bowl winner with ~70% confidence on the Sunday morning of; the favorite won.
For traders looking at our how to read a prediction market primer, the calibration property is what justifies treating prices as probabilities in the first place.
Where prediction markets beat polls
On binary, well-defined, source-resolvable events with deep liquidity, prediction markets have a structural advantage over polls:
Continuous updating. A poll is a snapshot from when the survey was fielded. A market price updates every time anyone trades. By the time a poll is published, the market has typically already absorbed any news the poll missed.
Skin in the game. Pollsters’ respondents have no financial stake in their answers. Market participants put real money on theirs. The set of people willing to bet typically aggregates more relevant information than the set of people willing to answer a survey.
Resolution-source clarity. A market resolves on a specific event with a specific source. A poll measures stated preference, which isn’t always the same as ultimate behavior.
The 2016 US presidential election is the most-cited counterexample — Brexit and Trump both surprised markets — but in both cases, the markets were less surprised than polls were. The Polymarket equivalent of 2016 (the Iowa Electronic Markets back then) had Clinton at ~75%; most polls had Clinton at 90%+. Markets called the upset more than polls did, just not strongly enough.
Where prediction markets struggle
Markets are not a universal forecasting tool. They underperform on:
Thin markets. A market with $50K of volume has none of the structural advantages above. A handful of traders can move the price; the headline number is fragile. Always check liquidity before treating a price as a forecast.
Ambiguous resolution. When the resolution criterion is vague, traders price both the underlying probability and the resolution-process risk. The price drifts away from a clean forecast. This is the failure mode our UMA oracle piece covers in detail.
Long horizons with no catalysts. A market resolving in 18 months with no news in between is mostly trader sentiment, not fundamentals. Calibration degrades as the horizon extends.
Politically charged or motivated trading. When a market becomes politically symbolic — Trump-related markets in 2020, certain election markets in 2024 — coordinated trading from politically motivated participants can pull the price away from the unbiased estimate. Researchers have documented this effect; it is small in aggregate but meaningful on specific high-profile markets.
Comparison: markets vs. major forecast aggregators
How do prediction markets compare to specific forecasting models?
vs. FiveThirtyEight (Silver’s models, 2008–2024 averaged): Roughly even. Markets typically slightly more confident on the favorite; aggregator models slightly more conservative. Both well-calibrated.
vs. RealClearPolling averages: Markets typically beat individual polls and roughly match polling averages on top-line presidential races. On individual House races, polling averages often beat markets due to limited liquidity.
vs. Metaculus: Metaculus and Polymarket usually agree within a few percentage points on overlapping questions. Where they diverge, Polymarket is biased toward shorter-horizon questions; Metaculus toward longer horizons and ambiguous resolution.
vs. internal models (Goldman, JPM, etc.): Prediction markets and Wall Street economic models give similar baseline numbers on macro questions like recession probability and Fed paths, with markets typically more responsive to news. The September 2024 yield-curve narrative is a good example: both markets and economists were tracking the inversion; markets repriced faster on each new data print.
For a live test of these comparisons, see our chance of recession and next Fed rate cut breakdowns, both of which compare Polymarket to FedWatch and to Wall Street estimates.
What the accuracy data does not tell you
Two limitations worth flagging:
Survivor bias. “Polymarket has 94% accuracy” includes only markets that resolved cleanly. Markets that were canceled, disputed, or otherwise not resolved are excluded from most accuracy reports. The true accuracy on a representative basket of markets is lower.
Calibration ≠ informativeness. A weather forecaster who says “70% chance of rain” every day in Seattle is probably well-calibrated. They are not particularly informative. Prediction-market accuracy depends both on calibration and on the market saying something useful — which usually requires a question with real disagreement.
How to use accuracy data in practice
For a reader trying to decide whether to take a prediction-market price seriously, the practical heuristics are:
- Is the market liquid? Daily volume above $100K is usually fine.
- Is the resolution rule clear? A named source and specific date is good; vague language is a yellow flag.
- Has a similar market resolved before? Precedent reduces resolution-process risk.
- Is the time horizon reasonable? Short-to-medium horizons (days to a few months) calibrate best. Longer horizons get noisier.
When all four check out, the prediction-market price is one of the best probabilistic forecasts you will find. When any of the four breaks down, treat the price as one input rather than the answer.
For new readers, the natural follow-ups are our how to read a prediction market primer and the Polymarket explained breakdown.
Common questions
How accurate are prediction markets?
On binary, well-defined, source-resolvable events with deep liquidity, prediction markets have historically been well-calibrated forecasters — meaning when the market says 70%, the event happens roughly 70% of the time across a large sample. Polymarket cites 94%+ accuracy on favorites in the final month before resolution.
Are prediction markets more accurate than polls?
On most major US elections since 2008, yes — markets have outperformed individual polls and roughly matched the best aggregators. The structural advantages are continuous updating, financial stake among participants, and clear resolution criteria.
When do prediction markets get it wrong?
Most often on thin markets (low liquidity), markets with ambiguous resolution criteria, long-horizon markets with no catalysts, and politically symbolic markets where coordinated trading pulls the price away from the unbiased estimate.
Did prediction markets call the 2016 election?
No, but they were less wrong than polls were. The Iowa Electronic Markets had Clinton at roughly 75% on election morning; most polls had Clinton at 90%+. Markets registered more uncertainty than the polling consensus did, which is the structural advantage in action — just not strong enough to flip the call.
How does Polymarket's accuracy compare to Kalshi's?
Both report comparable accuracy figures. Where they differ is on resolution-process risk: Kalshi uses traditional CFTC-regulated adjudication; Polymarket uses the UMA optimistic oracle. The accuracy of the underlying forecasts is similar; the resolution mechanics carry different risk profiles.
Should I bet based on prediction-market accuracy?
Aggregate accuracy data tells you the platform is well-calibrated on average. It does not tell you whether a specific market right now is correctly priced. Always check liquidity, resolution rule, time to resolution, and adjacent markets before treating a single price as a forecast.
