The honest answer: it depends on the type of market. Here is how calibration actually works — and where markets get it wrong.
If a market says 70%, and you ran that kind of event 100 times, how often did the 70¢ side win? A perfectly calibrated market would see that event happen exactly 70 times out of 100. If it happens 90 times, the market was underpricing it. If it happens 50 times, it was overpricing it.
Key concept: calibration is not the same as prediction. A prediction market does not predict the future — it aggregates the beliefs of participants who have skin in the game. The price is a probability estimate, not a certainty.
Good calibration means the market's probability estimates are systematically accurate across many outcomes. Bad calibration means the numbers are noise — they look like probabilities but don't behave like them over time.
These market categories consistently show better calibration: objective resolution sources, deep liquidity, and well-understood base rates.
Federal Reserve decisions, CPI releases, and jobs reports resolve against government-published data. Clear resolution source, no insider edge from within the market itself, and deep liquidity from institutional participants.
Historical base rates are well-established and markets are large with fast resolution. Sportsbook lines provide a calibration anchor — PM prices that diverge significantly from Vegas lines tend to get arbitraged back quickly.
NOAA and weather station data is objective, tamper-resistant, and fast to resolve. Nobody can influence whether it rains in a city on a given day — the resolution source is as clean as it gets.
Large national elections combine months of polling data with market-based conviction. Markets tend to show late-breaking responsiveness that polls miss, and the thick liquidity in major election markets reduces noise.
These conditions degrade market accuracy. Check for these before placing significant weight on a price.
Small markets with low volume are easy to move with a single large order. Price is noise, not signal — one participant with strong conviction (or bad intent) can set the displayed probability regardless of actual information.
Celebrity behavior markets, company announcement markets, and sports injury markets all have participants with non-public access. The MrBeast and OpenAI insider trading cases confirmed this attack vector is real.
Mention markets (did X happen before Y date?), vague resolution terms, and markets where operator discretion governs the outcome create genuine uncertainty about what the market is even pricing.
When an event is days away and almost no public information exists, markets cannot aggregate meaningful beliefs. Early prices on fast-breaking events can be nearly random until information surfaces.
Most confusion about PM accuracy starts with misreading what the displayed price actually represents.
When you see 65¢ on a contract, here is what it means — and what it does not mean
✅ Does mean
❌ Does not mean
The spread matters too. The displayed price is often the ask, not the true mid-price consensus. Learn how bid/ask spread affects the price you see →
The accuracy debate between markets and polling is more nuanced than either side admits.
The honest take: These are complementary, not competing. Polls measure stated opinion; markets measure financial conviction. Both fail in different ways. Polls fail when respondents lie or the turnout model is wrong. Markets fail when liquidity is thin or information is asymmetric. The most informed view of any event uses both.
