Why most prediction-market traders lose money

On 28 April 2026, Bloomberg reported that across roughly two million Polymarket addresses, traders were down about $131 million net — while automated participants were taking gains out of the same markets. The full piece is here: Most prediction market traders are losing money while bots rack up gains.

That figure is the one honest starting point for anyone thinking about trading these markets, and it deserves better than the two usual reactions. It is not evidence that prediction markets are rigged, and it is not evidence that people are stupid. It is the expected outcome of a specific cost structure meeting a specific set of habits. Every mechanism below is identifiable, and most of them are measurable on your own trading record.

The arithmetic that decides everything

A prediction-market contract pays 1 if the event resolves yes and 0 if it does not. Buy at price c, believe the true probability is p, and ignore costs for a second:

EV per contract = p x (1 - c) - (1 - p) x c = p - c

Your entire edge is the gap between your probability and the price. That is a clean, honest instrument — and it is also brutal, because the gap has to survive costs that are charged on the notional, not on the edge.

Taking liquidity on Polymarket costs roughly 3.5% of notional, the figure we publish next to our own calibration work in the public register. Put that against a typical edge. A trader who is genuinely 3 points better than the market — buying at 0.60 something that resolves 63% of the time — has an edge of 0.03 per contract and pays a cost of the same order of magnitude on the way in. Being right about the world is not sufficient. You have to be right by more than the cost of expressing it, on every trade, including the ones you would have skipped if you had run the arithmetic first.

Almost nobody runs the arithmetic first. This is not a small contributor to the $131 million; it is the largest single one, because it applies to every trade rather than to the bad ones.

The price is usually better than you think

The second mechanism is that the market's price is a genuinely strong forecast, and traders systematically underestimate it.

We measured this on our own terrain — five-minute BTC, ETH, SOL and XRP markets — by reading the market price partway through each window and comparing it with how the window actually resolved. The full table is published, all bands included. Two features of it matter here:

• —Across most of the price range, the market's price is close to the realised outcome. The gaps that exist are of the same order as the cost of taking liquidity.
• —In the low-price bands the gap is negative — those sides resolve less often than their price implies. Cheap contracts are, on our data, systematically worse than they look.

That last shape has been observed in betting markets for decades: long-shot prices are bid up beyond their realised frequency, because a contract that costs 6 cents and pays a dollar sells itself. It is the same trade as a lottery ticket, and it has the same expected value problem, made worse by the fee being a much larger fraction of a 6-cent entry.

The full numbers, with their coverage and the bands where our own signal is worse than the market, are on the public register. We publish the unflattering bands for the same reason this article exists: a calibration table with its losing rows removed is not a calibration table.

Your counterparty is probably a machine

Bloomberg's reporting is explicit that automated participants were on the profitable side. That is not a conspiracy; it is a description of what market making is.

An automated participant on these venues typically does three things a human cannot. It quotes both sides continuously and earns the spread from whoever crosses it. It manages inventory across correlated markets, so a position taken from you is hedged within seconds. And it updates its prices on new information faster than a human can read the headline that caused it.

The practical consequence: when you place a market order, you are trading against a participant whose business model is being paid for that trade. When you place a limit order and wait, you are doing the thing they are paid for. That single change in habit does not create an edge, but it stops handing one over on every fill.

What you are actually betting on

A prediction-market contract is not a bet on an event. It is a bet on a resolution rule applied to a named source at a stated time. The distinction is invisible until it costs you.

Markets resolve on the wording, and the wording contains all the ambiguity: what counts as an official announcement, which source is authoritative, what happens on a partial outcome, what happens if the deadline passes with no determination. Traders who "were right" and lost are almost always traders who read the title and not the rules.

There is also a cost most retail participants never price: capital is locked until resolution. On a market resolving in six months, your capital earns nothing while it waits. Against a risk-free alternative, a small positive edge on a long-dated contract can be a negative-return use of capital even when it wins.

Sizing kills the survivors

Suppose you clear all of the above: real edge, net of costs, on a market whose rules you have read. You can still go broke, and the mechanism is sizing.

For a binary contract bought at price c where you believe the probability is p, the Kelly fraction is:

f* = (p - c) / (1 - c)

Two things about that formula are worth internalising. First, it shrinks fast as c rises: the same 3-point edge justifies a much smaller stake on a 0.90 contract than on a 0.50 one, because you are risking 90 cents to make 10. Second, it is calculated from your estimate of p, and your estimate is wrong by some amount you do not know. Betting full Kelly on an overconfident probability is a reliable way to convert an edge into a drawdown you cannot sit through.

Binary payoffs have high variance by construction. A 60% strategy loses four in a row often enough that it will happen to you, and the only defence is a stake small enough that it does not matter.

The record-keeping problem

The final mechanism is why so few traders discover any of the above from their own experience: they never build the record that would show it.

What survives instead is a mixture of memory and screenshots — the resolved winners, the calls that felt clever, the positions people posted about because they were working. Losing positions are quietly not posted. The public record of prediction-market trading is a survivorship-filtered highlight reel, and traders calibrate their expectations against it.

The fix is unglamorous and effective: log every trade with the price you paid, the probability you believed, and the outcome. Then group your trades by the price you paid and compare your realised win rate with the average price of that bucket. That is your own calibration table, built exactly the way ours is. It answers the only question that matters — are you actually better than the price? — and it answers it in a way no amount of remembering can.

What actually helps

None of this makes prediction markets unbeatable. It makes them expensive, and it makes the requirements specific:

• —Price the cost before the view. If your edge is not clearly larger than the round-trip cost, the trade is negative expectancy no matter how confident you are.
• —Provide liquidity instead of taking it wherever the position allows waiting.
• —Read the resolution rules — the source, the deadline, the tie-breaks — before the title.
• —Be sceptical of cheap contracts. In our measurements the low-price bands resolve less often than they are priced.
• —Size from an edge you have measured, not from one you assume, and stay well under the fraction the formula gives you.
• —Trade fewer markets. Costs scale with the number of trades; edge does not.
• —Keep the record. Your own calibration table is the only feedback loop that is not survivorship-biased.

The $131 million is not a mystery. It is fees, spread, long-shot pricing, resolution surprises and position sizing, aggregated across two million addresses over time. Each of those is addressable, and none of them is addressed by having a better opinion about the world.

That is also why we publish measurements rather than forecasts. Our own register shows what our signal does and what it does not — including the price bands where the market is right and we are not, and the share of slots whose outcome we could never retrieve. A number you cannot check is worth exactly as much as a screenshot.

Related reading: how to read a level-2 order book and what funding rates say about positioning, on the crypto side of the same discipline.