Every few days a prediction market gets something "wrong," and the internet buries the whole idea. A 78% favourite crashes out. A contract sitting at 5% goes on to win. "So much for the wisdom of crowds." Then, a week later, the same people quote a market price as if it were gospel. Both reactions are wrong, and they're wrong in the same way — by mistaking a probability for a promise.
We benchmark our own AI forecaster against prediction markets every day, in public, so we've spent more time than most staring at exactly where these markets are sharp and where they crack. Here's the honest version — more useful than what either the cheerleaders or the critics will tell you.
First, the uncomfortable truth: they're very accurate
A prediction-market price is a probability. A contract that pays $1 if an event happens, trading at 63 cents, is the crowd saying "63%." And that number is astonishingly hard to beat.
The reason is money. A poll costs a respondent nothing to answer carelessly; a pundit pays no price for a confident miss. A market forces everyone to put cash behind their opinion, and cash is a truth serum. Research on market calibration keeps finding the same result: across all the contracts trading near 70 cents, the event happens roughly 70% of the time. Near 30 cents, about 30%. That property — saying 70% and being right 70% of the time — is called calibration, and it's the gold standard for a forecast. Most humans, most models, and almost all pundits fail it. Liquid markets usually pass.
So as a free, real-time forecast, the market price is one of the best instruments available. Start there.
But "accurate" doesn't mean "right every time"
Here's where most people trip. A market that prices a team at 70% and watches them lose has not failed. Seventy percent means the favourite loses three times in ten — by design. Judging a market on a single outcome is like judging a coin as "broken" because it landed tails.
The only honest test is the full record: across every event priced at 70%, did the favourite win about seventy times in a hundred? Over hundreds of markets, dozens of favourites fall, and every one becomes somebody's proof that "the odds were garbage." They weren't. You just watched one of the threes.
Internalise that and you've already beaten most of the people loudly dunking on prediction markets.
Where they genuinely break — two real failure modes
Markets aren't magic, and pretending otherwise is how you get fleeced. There are two distortions worth knowing cold.
**Favourite-longshot bias.** Across decades of betting and market data, one pattern refuses to die: longshots are overpriced. A contract at 4 cents — implying a 4% chance — tends to happen less than 4% of the time. Why? A few cents feels like a lottery ticket, and lottery tickets are fun. Multiply that by a crowd of emotionally invested participants and you get persistent, measurable overpricing at the tails. When you see a miracle priced at 5%, the sober read is usually "3%, plus vibes."
**Narrative gravity and thin liquidity.** Crowds love hosts, comeback stories and golden generations. Money doesn't fully launder sentiment out of a price — it just taxes it, and sometimes the tax is too low, so a team trades on its story rather than its squad. And liquidity is not wisdom: a deep market on a championship final is a serious instrument; a thin market on an obscure event at 3 a.m. is a rumour with a bid-ask spread. Same interface, very different epistemic weight.
How to actually read a market price
- **Read the volume before you read the price.** Thin market, weak signal. This one habit separates using the tool from being used by it. - **Treat the price as a baseline, and question it hardest when everyone agrees.** Consensus is where narrative pricing hides. - **Think in distributions, not outcomes.** "The market had it at 80% and it didn't happen" is one data point and tells you almost nothing. A hundred markets, scored, tell you almost everything. - **Respect the tails, but discount the miracles.** If it's priced like a lottery ticket, it probably is one.
The live test we run on ourselves
This isn't theory for us. Our AI's probabilities are scored against the prediction-market price on every World Cup match — each call locked before kickoff and hashed into Bitcoin so nothing can be backdated, then Brier-scored after the result. The board is public, losses and all.
The honest tally so far: the market is closer than our model most nights. Our model only edges ahead on average error because it refuses to round two "impossible" upsets down to zero — exactly the favourite-longshot blind spot above, working in our favour on the rare night it mattered. Neither side is an oracle. The disagreement between a disciplined model and an efficient crowd is the most informative number on the screen.
That's the real shape of it. A prediction market isn't a crystal ball and it isn't rigged. It's a very good, very cheap forecast with two well-documented blind spots. Learn to read it — baseline first, volume always, tails with suspicion — and you'll be right far more often than the people yelling that the odds are broken every time a favourite falls.
*Educational content about forecasting and prediction markets — not betting or financial advice.*