How efficient are prediction markets? Calibration across 9.46 million trades

A prediction market is efficient when its prices work as probabilities: a contract at 30¢ should come true about 30% of the time. Plot price against how often each price band resolved true, and a perfectly efficient market traces the diagonal.

0xinsider's analytics pipeline tracks 1,236,804 resolved Polymarket markets and 44.1 million resolved trader-position rows. Binning every resolved position by the price the trader paid, and counting how often each band won, gives a curve over about 9.46 million trades. At the ends it sits almost on the diagonal: a contract bought near 5¢ resolved true 3.5% of the time, and one bought near 95¢ resolved true 98.4% of the time.

So the price is a usable probability, and it is closest to exact at the ends. The curve leaves the diagonal in the middle.

A position bought around 45¢ resolved true 52.6% of the time, 7.6 points above its price. A 55¢ position won 63.3% (+8.3 points), a 65¢ position 73.2% (+8.2), and a 75¢ position 82.2% (+7.2). The gap shrinks toward the top: +5.6 points at 85¢ and +3.4 at 95¢.

The market is sharpest where the crowd agrees and softest near a coin flip. A 55¢ contract is where disagreement, thin liquidity, and slower information leave room, and where a trader with better calibration has something to capture.

Two things inflate that gap. The curve is built from positions traders chose to hold to resolution, and traders may hold on to mid-range bets they believed were underpriced. And the price we measure against is the trader's average fill, not the market's mid-quote at the moment of the trade. Read the middle of the curve as where edge appeared in held positions, not as proof that any 55¢ contract is underpriced.

The core metric is calibration edge: the average of outcome minus entry price across a trader's resolved markets, where outcome is 1 for a win and 0 for a loss. A trader who keeps buying winners below their realized frequency posts a positive number; one who overpays for favorites posts a negative one. In an efficient market the baseline is zero: a 95¢ favorite pays +5¢ when it wins (95% of the time) and -95¢ when it loses (5%), which nets to zero over a real sample. The companion metric is the Brier score, the mean squared error between price and outcome, where lower is better.

Three rules keep the sample clean. Voided and refunded markets are dropped, since they have no real outcome. Multi-leg positions collapse to the trader's net side per market, so a hedged account does not double-count. And the metric only computes once a trader has at least 5 resolved markets, below which one lucky call would dominate. Across the population, the Brier score averages 0.149 with a median of 0.154, which is what informative but imperfect prices produce.

Two limits, stated so a third party can replicate and critique this. First, entry price is the trader's average fill, not the market mid at trade time, so the metric mixes forecasting skill with execution and timing; separating them against order-book mid quotes is a next step this version does not claim. Second, every number here is Polymarket-only. Polymarket positions carry public wallet identity, which makes per-trader calibration possible; a venue whose positions are not public cannot be studied this way.

Across 8,705 Polymarket traders with at least 5 resolved markets, the average calibration edge is +5.5¢ and the median +2.2¢. 70% of them (6,096) post a positive edge and 30% (2,609) a negative one. The typical trader in this sample is mildly sharp, and a long tail is sharper, with the selection caveat above applying to both.

The 10th percentile trader sits at -3.9¢, the 25th at -0.5¢, the 75th at +9.0¢, and the 90th at +23.2¢, so the top decile bought positions that resolved true more than 20 points above what they paid. A 5-market floor is low, though, so the top of the distribution mixes skill with luck, and the high percentiles are the least stable numbers here.

Win rate rewards anyone who buys favorites: a trader who only backs 90¢ locks can post a high win rate while making almost nothing, because the price already reflected the outcome. Calibration edge rewards paying less than the eventual frequency. In the 0xinsider grade it is one input to the roughly 5% skill tie-breaker; realized profit drives the rest.

For a trader, the curve shows where the gap was biggest: 45¢ to 75¢, at 7.2 to 8.3 points. Near 95¢ it was 3.4 points, and near 5¢ the price ran 1.5 points above the outcome rate. The selection caveat applies across the curve, so check your own record first: across your resolved markets, does the outcome beat your entry price, or are you paying up for favorites?

For an analyst, the headline is that Polymarket prices behave like probabilities across 9.46 million resolved trades, with one located deviation instead of scattered noise. The method is reproducible: bin price against realized frequency over public resolution data, and compute the per-trader distribution from the same source.

Wallet profiles show calibration edge where it has been computed, beside grade and P&L, so you can check one trader against the aggregate here.