Prediction markets · calibration

How much do Polymarket prices know, and when?

Shmuel San · October 2026

Jason Furman's post after the 2024 US election showed contract prices lining up with win rates on the 45° line. This page rebuilds that chart from Polymarket's own data, then breaks it down by time before the event, type of event, trading volume and number of traders, with market-level regressions. The second page looks at individual traders' profits, at prices after large orders, at trading that looks like pushing the price and at whether knowing who trades beats the price; the third applies all of it to the 2026 Israeli election markets, and the fourth looks at who keeps trading and how traders behave.

61,518 resolved markets 15,492 events Oct 2023–Oct 2026 volume ≥ $1,000

What the data show

Prices are traded prices. For every market and every day, the price used is the last price at which someone actually traded, from Polymarket's trade records. Polymarket's own price-history series is a quote that, in markets with an empty order book, sits near 50¢ whatever the odds; see "Quoted prices against traded prices" below.

Furman's picture holds, and the longshot bias disappears. Pooling 586,610 market-days, win rates sit on the 45° line. With traded prices, contracts under 10¢ win within half a cent of their price at every horizon. The calibration slope is 0.98–0.99 from 1 day to 1 week out and falls to about 0.92 at 2–3 months, where favourites above 90¢ are overpriced by about 5¢.
Information arrives steadily over the last three months. Following the same 10,059 markets every day, Brier skill (the share of outcome uncertainty the price resolves) rises from 0.46 90 days before the event to 0.64 at 30 days, 0.76 at a week and 0.82 at a day.
The number of traders predicts signal; lifetime volume barely does. A week out, Brier skill rises from 0.38 in the bottom quintile of wallets that had traded so far to 0.70 in the top, and the gradient holds within every event type, from the bottom to the top trader-count tercile (sports 0.39 → 0.67, politics 0.44 → 0.75, crypto 0.35 → 0.70). By lifetime volume, skill is flat (0.53–0.57). In the regressions, a week out each doubling of traders adds about 0.08 of Brier skill holding type, dollar volume, market age, staleness and year fixed (0.05 a day out); a month out the effect is not distinguishable from zero once dollar volume is controlled. Knowing which wallets trade adds only about 1% to forecasts beyond the price a day out (traders page, section 7). Thin markets are also too extreme: their calibration slope is about 0.86–0.91.
"Will X say / tweet …" markets know the least, even when crowded. With all controls they trail elections by about 0.30 of Brier skill a week and a day out (0.18, less precisely estimated, a month out). With the most traders they reach only 0.37 a week out, and their prices are too extreme (slope 0.87). This is the one type where the person named in the question controls the outcome. A large order there moves the price 3.1¢, more than in any other type, and the move is gone a week later (traders page).
Prices overreact. Holding today's price fixed, last week's price move predicts the outcome with a negative sign: about 15% of a week's move reverses by resolution. The reversal is larger in thin markets (about 23%) and, a day before the event, in mentions markets (about 23%). That is the pattern noise traders, or someone pushing the price, would leave. The traders page measures how much trading looks like pushing the price (a few percent of large-order dollars), and finds that prices after large orders keep moving about as often as they come back (momentum vs. reversal).
Elections are informed early; other types catch up late. The gap between elections and crypto, geopolitics, politics, tech and economy markets is widest far from the event and closes as it approaches: elections already sit at 0.59 of skill a month out, but a day out these types match or beat elections (0.64).

Furman's chart, rebuilt

Each dot is one of 50 price bins, pooling 586,610 market-days: every market's last traded price once a day over its last two weeks. Like Furman's chart, it lines up on the diagonal.

This is the easiest version of the test. Most market-days are priced near 0 or 1 just before resolution, and repeated days of the same market count as separate observations. The sections below take one price per market at a fixed distance from the event instead.

1 · Time before the event

Each panel takes one price per market, the last trade a fixed time before the event, and plots the share of markets that resolved Yes, in 20 price bins with 95% intervals. The dashed line is perfect calibration.

Signal: Brier skill by time before the event

"Same markets" follows the 10,059 markets priced at every horizon from three months to a day, so its rise shows information arriving. "All markets" uses every market priced at each horizon, so it mixes in short-lived markets close to the event.

Bias: calibration slope by time before the event

1 (dashed) is calibrated. Below 1 means prices are too extreme for what they know.

Longshot gap by time before the event (contracts under 10¢)

Price minus win rate, in dollars. 0.01 means longshots are overpriced by one cent.

Calibration by time before the event

3 months before10,059 markets

2 months before13,485 markets

1 month before21,415 markets

3 weeks before26,285 markets

2 weeks before31,500 markets

1 week before39,786 markets

3 days before51,220 markets

2 days before55,415 markets

1 day before61,518 markets

Summary by horizon, all markets observed at that horizon

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
3 months	10,059	2,053	0.12	0.4550.420–0.484	0.4560.424–0.480	0.9280.917–0.936	0.920.89–0.95	0.0030.001–0.006	0.0550.013–0.098
2 months	13,485	2,927	0.13	0.4840.458–0.510	0.4790.453–0.502	0.9360.931–0.943	0.930.90–0.95	0.0040.002–0.007	0.0510.020–0.084
1 month	21,415	4,941	0.15	0.5040.481–0.527	0.4970.476–0.516	0.9390.934–0.944	0.940.92–0.96	0.002-0.000–0.004	0.0300.010–0.053
3 weeks	26,285	6,073	0.16	0.5140.495–0.533	0.5060.487–0.523	0.9400.935–0.944	0.960.94–0.98	0.001-0.000–0.003	0.0180.007–0.033
2 weeks	31,500	7,288	0.17	0.5300.513–0.546	0.5160.499–0.533	0.9410.936–0.945	0.980.96–0.99	0.000-0.002–0.002	0.0230.009–0.036
1 week	39,786	9,313	0.18	0.5530.541–0.568	0.5340.522–0.548	0.9450.942–0.948	0.980.98–1.00	-0.001-0.003–0.000	0.0120.003–0.020
3 days	51,220	12,020	0.20	0.5570.547–0.569	0.5350.525–0.546	0.9450.942–0.948	0.980.97–0.99	-0.001-0.003–0.000	0.0210.014–0.029
2 days	55,415	13,323	0.21	0.5580.544–0.567	0.5360.525–0.546	0.9450.942–0.947	0.980.97–0.99	-0.002-0.003–-0.000	0.0150.009–0.021
1 day	61,518	15,492	0.23	0.5650.553–0.574	0.5430.532–0.552	0.9460.943–0.948	0.990.98–0.99	-0.001-0.003–-0.000	0.0120.007–0.019

Below each estimate: 95% interval from a bootstrap that resamples whole events. The sample differs by horizon, because only longer-lived markets have a price 2–3 months out.

Same markets at every horizon (10,059 markets priced from 3 months to 1 day before)

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
3 months	10,059	2,053	0.12	0.4550.418–0.483	0.4560.422–0.481	0.9280.918–0.936	0.920.89–0.95	0.0030.001–0.006	0.0550.023–0.095
2 months	10,059	2,053	0.12	0.5350.508–0.569	0.5260.496–0.556	0.9480.941–0.956	0.960.93–0.99	0.0050.002–0.007	0.0410.004–0.082
1 month	10,059	2,053	0.12	0.6370.610–0.667	0.6300.606–0.656	0.9700.965–0.975	1.000.98–1.02	0.0040.002–0.006	0.002-0.014–0.019
3 weeks	10,059	2,053	0.12	0.6790.653–0.710	0.6680.644–0.696	0.9770.972–0.981	1.010.99–1.03	0.0030.001–0.005	-0.001-0.015–0.014
2 weeks	10,059	2,053	0.12	0.7200.695–0.741	0.7050.683–0.724	0.9820.978–0.985	1.021.00–1.04	0.0030.001–0.005	0.001-0.013–0.019
1 week	10,059	2,053	0.12	0.7560.735–0.780	0.7380.719–0.761	0.9850.981–0.989	1.021.01–1.03	0.0030.001–0.004	-0.006-0.016–0.007
3 days	10,059	2,053	0.12	0.8020.780–0.822	0.7810.763–0.799	0.9900.987–0.992	1.021.01–1.04	0.0030.002–0.005	-0.004-0.013–0.006
2 days	10,059	2,053	0.12	0.8050.782–0.827	0.7860.767–0.806	0.9910.988–0.993	1.011.00–1.03	0.0030.001–0.004	-0.005-0.014–0.005
1 day	10,059	2,053	0.12	0.8170.799–0.838	0.7970.780–0.816	0.9910.989–0.994	1.011.00–1.03	0.0020.000–0.003	-0.005-0.012–0.004

Holding the sample fixed removes the composition effect, so this is the cleaner read of how information accumulates as the event approaches.

Calibration is close to perfect up to a week out and slips at 1–3 months, where long-dated favourites are overpriced. What changes much more is how far prices have moved toward 0 or 1. Across all markets, skill falls only gently with the horizon (0.57 a day out, 0.46 at three months), because the sample changes: short horizons add many short-lived sports markets that are hard to call before they are played. Holding the market set fixed shows the steeper rise, from 0.46 to 0.82.

2 · Type of event

Event types come from Polymarket's tags, with "will X say / mention / tweet …" questions pulled out first, because the person named can directly decide the outcome. Calibration slopes are close to 1 for most types; the exceptions are mentions markets (0.87 a week out, 0.80 a month out) and, at a month, geopolitics and "other" (0.80–0.84).

Signal: Brier skill by event type, 1 day, 1 week and 1 month before

Sorted by skill a week before. Groups with fewer than 100 markets at a horizon are left out.

Signal by time before the event, for each event type

Elections

Tech/Science

Politics

Sports

Culture

Economy/Fed

Crypto

Other

Geopolitics

Esports

Mentions/tweets

Same scale in every panel; shaded band is the 95% interval. These mix market sets across horizons, as in "All markets" above.

Bias: calibration slope by event type

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	0.36	0.42	0.49	0.51	0.54	0.55	0.53	0.52	0.51
Elections	0.60	0.62	0.59	0.59	0.60	0.62	0.63	0.63	0.64
Culture	0.44	0.47	0.48	0.53	0.52	0.55	0.58	0.60	0.65
Politics	0.41	0.46	0.50	0.52	0.55	0.58	0.57	0.61	0.66
Economy/Fed	0.48	0.44	0.49	0.41	0.46	0.52	0.58	0.60	0.64
Geopolitics	0.37	0.38	0.31	0.36	0.41	0.48	0.57	0.60	0.68
Crypto	0.19	0.29	0.43	0.48	0.54	0.52	0.58	0.65	0.71
Mentions/tweets	·	·	0.22	0.11	0.13	0.25	0.29	0.29	0.30
Tech/Science	0.41	0.44	0.52	0.59	0.61	0.61	0.66	0.68	0.74
Esports	·	0.17	0.42	0.47	0.44	0.45	0.45	0.43	0.49
Other	·	·	0.29	0.45	0.45	0.48	0.53	0.57	0.58

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	0.39	0.44	0.51	0.52	0.54	0.55	0.52	0.52	0.51
Elections	0.58	0.60	0.57	0.57	0.57	0.60	0.61	0.60	0.61
Culture	0.44	0.46	0.47	0.51	0.50	0.53	0.56	0.58	0.62
Politics	0.38	0.43	0.48	0.49	0.51	0.54	0.54	0.58	0.63
Economy/Fed	0.48	0.42	0.46	0.40	0.43	0.47	0.52	0.55	0.58
Geopolitics	0.34	0.38	0.32	0.37	0.41	0.47	0.54	0.57	0.64
Crypto	0.21	0.30	0.42	0.46	0.50	0.50	0.56	0.62	0.68
Mentions/tweets	·	·	0.17	0.09	0.11	0.21	0.25	0.24	0.25
Tech/Science	0.40	0.40	0.49	0.55	0.56	0.56	0.61	0.64	0.70
Esports	·	0.12	0.41	0.46	0.43	0.43	0.43	0.41	0.47
Other	·	·	0.23	0.41	0.41	0.44	0.47	0.51	0.52

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	0.911	0.930	0.948	0.949	0.951	0.952	0.946	0.944	0.941
Elections	0.954	0.961	0.951	0.953	0.952	0.959	0.961	0.960	0.962
Culture	0.930	0.929	0.927	0.942	0.936	0.942	0.948	0.954	0.962
Politics	0.911	0.930	0.930	0.925	0.937	0.946	0.944	0.955	0.965
Economy/Fed	0.949	0.923	0.923	0.896	0.913	0.925	0.938	0.944	0.954
Geopolitics	0.910	0.912	0.888	0.900	0.913	0.933	0.950	0.956	0.969
Crypto	0.842	0.888	0.914	0.930	0.936	0.934	0.953	0.964	0.973
Mentions/tweets	·	·	0.826	0.804	0.795	0.805	0.819	0.816	0.819
Tech/Science	0.911	0.891	0.929	0.948	0.947	0.945	0.960	0.966	0.978
Esports	·	0.832	0.898	0.925	0.911	0.907	0.919	0.912	0.927
Other	·	·	0.879	0.909	0.900	0.914	0.925	0.937	0.942

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	0.85	0.88	0.92	0.94	0.97	0.97	0.95	0.95	0.97
Elections	1.03	1.01	0.99	1.00	1.00	1.00	1.00	0.99	0.99
Culture	0.92	0.94	0.95	0.99	0.97	0.98	0.99	0.99	1.01
Politics	0.90	0.92	0.98	1.00	1.04	1.05	1.02	1.03	1.03
Economy/Fed	0.95	0.96	1.03	0.99	0.97	0.97	1.00	1.00	1.02
Geopolitics	0.90	0.88	0.84	0.95	1.03	1.08	1.08	1.06	1.05
Crypto	0.74	0.81	0.91	0.94	1.02	1.00	1.00	1.04	1.04
Mentions/tweets	·	·	0.80	0.75	0.74	0.87	0.93	0.92	0.92
Tech/Science	0.96	0.92	0.97	1.06	1.05	1.03	1.04	1.03	1.04
Esports	·	0.67	0.93	1.01	0.96	0.96	0.93	0.91	0.94
Other	·	·	0.80	0.99	1.00	1.06	1.08	1.10	1.08

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	-0.002	0.000	0.001	0.000	0.001	0.001	-0.001	-0.000	0.000
Elections	0.006	0.004	-0.003	-0.002	-0.003	-0.001	-0.000	-0.002	-0.002
Culture	0.005	0.005	0.001	0.006	0.000	0.002	0.000	-0.002	-0.002
Politics	0.020	0.017	0.012	0.003	0.001	-0.001	-0.006	-0.003	-0.003
Economy/Fed	0.031	0.020	0.011	-0.011	-0.010	-0.017	-0.009	-0.010	-0.009
Geopolitics	0.037	0.026	0.016	0.015	0.009	0.003	-0.007	-0.010	-0.007
Crypto	0.011	0.033	0.009	0.013	0.005	-0.001	0.000	-0.000	-0.003
Mentions/tweets	·	·	0.017	0.027	0.022	-0.013	0.012	0.013	0.008
Tech/Science	-0.003	0.007	0.008	0.011	0.004	-0.010	-0.001	0.000	-0.002
Esports	·	-0.014	-0.005	-0.001	-0.005	-0.008	-0.001	-0.000	0.004
Other	·	·	0.034	0.013	-0.016	-0.027	-0.029	-0.032	-0.027

	3 months	2 months	1 month	3 weeks	2 weeks	1 week	3 days	2 days	1 day
Sports	0.36	0.43	0.57	0.62	0.68	0.73	0.81	0.81	0.83
Elections	0.60	0.65	0.70	0.73	0.76	0.78	0.80	0.80	0.80
Culture	0.44	0.54	0.70	0.73	0.76	0.79	0.81	0.82	0.83
Politics	0.41	0.50	0.67	0.67	0.70	0.69	0.72	0.73	0.79
Economy/Fed	0.48	0.59	0.71	0.73	0.78	0.80	0.85	0.86	0.86
Geopolitics	0.37	0.55	0.61	0.66	0.68	0.75	0.74	0.80	0.83
Crypto	0.19	0.37	0.46	0.55	0.60	0.65	0.73	0.74	0.76
Mentions/tweets	·	·	·	·	·	·	·	·	·
Tech/Science	0.41	0.58	0.72	0.77	0.77	0.88	0.89	0.89	0.90
Esports	·	·	·	·	·	·	·	·	·
Other	·	·	·	·	·	·	·	·	·

Darker means more of the metric. Cells with fewer than 100 markets are blank. Hover a cell for its 95% interval. Longshot gap is the mean price minus the win rate for contracts under 10¢; positive means longshots are overpriced.

Calibration by event type, 1 week before

Sports15,098 markets

Elections8,431 markets

Culture5,185 markets

Politics2,062 markets

Economy/Fed1,950 markets

Geopolitics1,875 markets

Crypto1,447 markets

Mentions/tweets1,371 markets

Tech/Science1,036 markets

Esports972 markets

Other359 markets

Summary by event type, 1 week before

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Sports	15,098	1,966	0.14	0.5540.523–0.582	0.5470.521–0.572	0.9520.947–0.957	0.970.95–0.99	0.001-0.001–0.003	0.0270.009–0.047
Elections	8,431	1,832	0.19	0.6190.594–0.643	0.5990.574–0.620	0.9590.953–0.964	1.000.98–1.02	-0.001-0.004–0.002	-0.002-0.014–0.011
Culture	5,185	1,028	0.17	0.5510.506–0.597	0.5270.486–0.565	0.9420.931–0.951	0.980.94–1.01	0.002-0.003–0.006	0.014-0.012–0.042
Politics	2,062	1,061	0.20	0.5750.527–0.622	0.5440.497–0.582	0.9460.934–0.956	1.051.00–1.10	-0.001-0.008–0.005	0.013-0.022–0.055
Economy/Fed	1,950	941	0.31	0.5200.474–0.567	0.4690.420–0.520	0.9250.911–0.938	0.970.93–1.00	-0.017-0.030–-0.005	0.0460.004–0.090
Geopolitics	1,875	1,061	0.20	0.4790.423–0.525	0.4680.417–0.513	0.9330.917–0.948	1.081.03–1.14	0.003-0.005–0.012	0.004-0.035–0.052
Crypto	1,447	631	0.18	0.5160.446–0.573	0.5040.431–0.560	0.9340.912–0.951	1.000.93–1.06	-0.001-0.012–0.009	-0.005-0.036–0.052
Mentions/tweets	1,371	158	0.38	0.2480.159–0.327	0.2060.127–0.286	0.8050.767–0.845	0.870.78–0.95	-0.013-0.053–0.021	0.011-0.029–0.077
Tech/Science	1,036	341	0.26	0.6130.551–0.676	0.5620.502–0.626	0.9450.928–0.960	1.030.98–1.07	-0.010-0.022–0.002	-0.027-0.040–-0.000
Esports	972	175	0.16	0.4530.364–0.548	0.4290.326–0.518	0.9070.864–0.938	0.960.88–1.03	-0.008-0.019–0.004	-0.006-0.035–0.058
Other	359	125	0.29	0.4840.372–0.574	0.4400.332–0.535	0.9140.871–0.947	1.060.96–1.18	-0.027-0.068–0.005	-0.014

3 · Trading volume

Volume is each market's lifetime traded volume, in quintiles. With traded prices, it predicts signal only weakly: a day out, skill is 0.53 in the bottom quintile and 0.62 in the top; a week out the quintiles are within 0.04 of each other. Thin markets are still too extreme far from the event (slope 0.82 a month out in the bottom quintile). With quoted prices instead, the volume gradient looks much steeper, but most of that comes from meaningless quotes in untraded markets.

Signal: Brier skill by lifetime volume quintile, 1 day, 1 week and 1 month before

Bias: calibration slope by volume quintile

Far from the event, the thinnest markets are priced much too extreme.

Signal: Brier skill by volume tercile within each event type, 1 week before

Terciles are computed within each event type, so each row compares thin and thick markets on the same kind of question.

Brier skill by volume quintile

AUC by volume quintile

Calibration by volume quintile, 1 week before

Q1 · $1,000–$2,9766,614 markets

Q2 · $2,976–$8,3427,376 markets

Q3 · $8,342–$23,7617,837 markets

Q4 · $23,761–$96,7468,116 markets

Q5 · $96,746–$1,531,479,2859,843 markets

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Q1 · $1,000–$2,976	6,614	2,116	0.12	0.5270.486–0.570	0.5230.489–0.557	0.9490.941–0.957	0.920.88–0.95	0.003-0.001–0.006	0.0490.020–0.081
Q2 · $2,976–$8,342	7,376	2,779	0.16	0.5620.523–0.586	0.5410.508–0.567	0.9470.939–0.954	0.970.94–1.00	-0.000-0.004–0.004	0.014-0.004–0.039
Q3 · $8,342–$23,761	7,837	3,251	0.19	0.5380.510–0.564	0.5160.489–0.542	0.9400.932–0.947	0.970.94–0.99	-0.000-0.005–0.003	0.0200.001–0.043
Q4 · $23,761–$96,746	8,116	3,654	0.21	0.5420.519–0.565	0.5150.493–0.538	0.9370.931–0.944	0.980.96–1.00	-0.002-0.006–0.001	0.003-0.014–0.021
Q5 · $96,746–$1,531,479,285	9,843	3,438	0.22	0.5660.544–0.591	0.5480.529–0.573	0.9490.944–0.955	1.031.01–1.05	-0.004-0.007–-0.001	-0.002-0.014–0.010

Volume within event type: Brier skill 1 week before, by volume tercile within each type

	Low volume	Mid volume	High volume
Sports	0.52	0.55	0.56
Elections	0.69	0.57	0.60
Culture	0.52	0.53	0.58
Politics	0.55	0.58	0.56
Economy/Fed	0.48	0.53	0.54
Geopolitics	0.55	0.45	0.42
Crypto	0.34	0.60	0.51
Mentions/tweets	0.25	0.25	0.24
Tech/Science	0.57	0.59	0.65
Esports	0.53	0.54	0.32
Other	0.39	0.43	0.62

Terciles are computed within each event type, so this compares thin and thick markets on the same kind of question.

4 · Traders and volume before the horizon

Here markets are sorted by how many distinct wallets, and how many dollars, had traded by 1 week and by 1 day before the event. Unlike lifetime volume, these measure only activity the price at that moment could reflect, and wallet counts are harder to inflate with wash trading than dollar volume.

This is a correlation. More traders may make prices sharper, or traders may crowd into questions where information is easier to find. The regressions in section 6 hold the event type, market age and other controls fixed, but cannot rule out the second story.

Counts come from up to 10,000 trades before each horizon per market, from 61,431 markets. 4% of market-horizons hit that cap; their counts are lower bounds and they all sit in the top quintile.

1 week before: by distinct wallets that had traded by then

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Q1 · 2–15 wallets	7,865	2,140	0.18	0.3830.341–0.418	0.3580.322–0.392	0.8940.881–0.904	0.870.83–0.91	-0.009-0.015–-0.004	0.1080.072–0.142
Q2 · 15–34 wallets	7,864	2,628	0.16	0.5280.496–0.554	0.5210.490–0.545	0.9440.935–0.950	0.970.95–1.00	0.000-0.003–0.003	0.020-0.008–0.044
Q3 · 34–68 wallets	7,865	2,862	0.17	0.5690.544–0.597	0.5550.531–0.578	0.9510.945–0.957	1.010.99–1.03	-0.000-0.004–0.003	-0.006-0.021–0.014
Q4 · 68–178 wallets	7,864	3,425	0.20	0.5780.558–0.601	0.5590.537–0.582	0.9490.945–0.956	1.000.99–1.02	0.002-0.002–0.005	-0.002-0.015–0.012
Q5 · 178–4,766 wallets	7,865	3,231	0.20	0.6960.676–0.718	0.6740.656–0.693	0.9750.972–0.979	1.041.03–1.06	0.0050.002–0.007	-0.013-0.021–-0.002

1 week before: by dollar volume traded by then

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Q1 · $0–$407	7,865	2,144	0.17	0.3770.340–0.410	0.3570.324–0.391	0.8950.884–0.907	0.860.83–0.90	-0.012-0.018–-0.008	0.1100.070–0.156
Q2 · $408–$1,814	7,864	2,633	0.15	0.5670.541–0.594	0.5610.538–0.586	0.9550.949–0.961	1.000.98–1.03	0.002-0.001–0.004	0.006-0.015–0.023
Q3 · $1,814–$7,640	7,865	3,190	0.19	0.5310.503–0.559	0.5110.485–0.537	0.9380.930–0.944	0.980.95–1.01	-0.001-0.005–0.004	0.006-0.013–0.027
Q4 · $7,641–$41,800	7,864	3,592	0.21	0.5900.565–0.611	0.5660.541–0.586	0.9510.945–0.956	1.010.99–1.02	0.003-0.001–0.006	0.001-0.012–0.015
Q5 · $41,800–$25,815,589	7,865	3,177	0.19	0.6910.670–0.716	0.6710.651–0.694	0.9750.971–0.979	1.041.03–1.06	0.0050.003–0.007	-0.014-0.024–-0.004

1 week before: Brier skill by traders and by dollar volume, quintiles measured up to the horizon

1 week before: Brier skill by trader-count tercile within event type

1 week before: the same, as a table

	Fewest traders	Middle	Most traders
Sports	0.39	0.56	0.67
Elections	0.52	0.61	0.68
Culture	0.43	0.59	0.62
Politics	0.44	0.49	0.75
Economy/Fed	0.41	0.50	0.64
Geopolitics	0.44	0.47	0.60
Crypto	0.35	0.49	0.70
Mentions/tweets	0.12	0.24	0.37
Tech/Science	0.36	0.69	0.77
Esports	0.38	0.51	0.48
Other	0.31	0.54	0.61

1 day before: by distinct wallets that had traded by then

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Q1 · 2–16 wallets	12,287	3,956	0.24	0.3740.348–0.396	0.3510.325–0.372	0.8880.878–0.896	0.880.85–0.90	-0.010-0.015–-0.006	0.1320.099–0.163
Q2 · 16–36 wallets	12,286	4,267	0.21	0.5600.535–0.579	0.5430.524–0.560	0.9470.941–0.951	0.990.97–1.01	0.001-0.002–0.004	0.003-0.010–0.019
Q3 · 36–70 wallets	12,286	4,614	0.20	0.5830.563–0.598	0.5690.549–0.583	0.9540.949–0.957	1.000.99–1.02	0.001-0.001–0.003	-0.011-0.020–0.001
Q4 · 70–173 wallets	12,286	5,321	0.24	0.5920.576–0.611	0.5670.550–0.584	0.9510.947–0.955	1.010.99–1.02	-0.000-0.003–0.002	0.001-0.009–0.011
Q5 · 173–4,845 wallets	12,286	5,154	0.25	0.7120.696–0.727	0.6830.667–0.698	0.9740.971–0.977	1.021.01–1.03	0.001-0.002–0.003	-0.011-0.017–-0.005

1 day before: by dollar volume traded by then

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
Q1 · $0–$564	12,287	3,911	0.24	0.3750.348–0.398	0.3500.325–0.371	0.8870.878–0.895	0.890.86–0.91	-0.014-0.018–-0.009	0.1280.093–0.169
Q2 · $564–$2,067	12,286	4,621	0.21	0.5460.524–0.566	0.5320.511–0.548	0.9440.939–0.948	0.990.97–1.01	-0.000-0.003–0.002	0.005-0.009–0.020
Q3 · $2,067–$7,496	12,286	5,133	0.22	0.5680.548–0.585	0.5480.528–0.565	0.9470.942–0.952	1.000.98–1.02	0.002-0.000–0.005	0.004-0.006–0.017
Q4 · $7,497–$38,006	12,286	5,338	0.22	0.6200.600–0.639	0.5950.577–0.613	0.9580.953–0.962	1.000.99–1.02	0.001-0.002–0.004	-0.005-0.014–0.006
Q5 · $38,022–$25,835,899	12,286	4,949	0.24	0.7210.703–0.737	0.6940.676–0.710	0.9770.974–0.980	1.031.02–1.04	0.0020.000–0.005	-0.010-0.016–-0.003

1 day before: Brier skill by traders and by dollar volume, quintiles measured up to the horizon

1 day before: Brier skill by trader-count tercile within event type

1 day before: the same, as a table

	Fewest traders	Middle	Most traders
Sports	0.37	0.54	0.61
Elections	0.55	0.63	0.69
Culture	0.56	0.69	0.69
Mentions/tweets	0.19	0.31	0.37
Politics	0.59	0.60	0.78
Geopolitics	0.63	0.67	0.76
Economy/Fed	0.58	0.60	0.74
Crypto	0.66	0.69	0.78
Esports	0.40	0.59	0.48
Tech/Science	0.61	0.80	0.81
Other	0.42	0.57	0.74

5 · Every day before the event

The same measures for each of the 90 days before the event, using each market's last traded price on that day. The horizontal axis is on a log scale so the final week is readable.

Signal: Brier skill, each of the last 90 days

"Same markets" follows the 10,059 markets that had traded at least once 90 days before the event. Shaded bands: 95% intervals, resampling events.

Bias: calibration slope, each of the last 90 days

By traders so far (terciles, recomputed each day)

By lifetime volume quintile

By event type

Elections

Tech/Science

Politics

Sports

Culture

Economy/Fed

Crypto

Other

Geopolitics

Esports

Mentions/tweets

Same scale in every panel; days with fewer than 100 markets are left out.

6 · Regressions: what predicts how much a market's price knows

Each observation is one market at a fixed number of days before the event. The dependent variable in tables A and E is the market's contribution to Brier skill: how much its price beat the event type's base rate, scaled so that its average within a type equals that type's Brier skill. Coefficients therefore read in Brier-skill units: 0.10 means 0.10 more of the outcome uncertainty resolved. Standard errors are clustered by event.

A. Traders predict skill a week and a day out: with all controls, +0.11 per log point a week out (+0.08 per doubling) and +0.08 a day out; a month out the coefficient is 0.03 (s.e. 0.03) once dollar volume enters. Dollars traded so far add to that. Given activity so far, higher lifetime volume predicts less skill: questions that stay contested keep attracting trading. Older and staler markets know less. Markets that resolved early ("by date X" questions) know more, partly by construction.
B. Outcome on price. The pooled slope is 0.99. It is steeper with more traders (+0.04 per log point) and flatter in thin markets (0.91) and mentions markets (0.87), meaning prices there are too extreme.
C. How far each price moved toward the outcome over the last month. Tech, crypto and geopolitics markets (and the small "other" group) learn the most in their final month; older markets learn less.
D. Overreaction: last week's price change predicts the outcome with coefficient −0.15 a week out and −0.07 a day out. In thin markets it is about −0.23 a week out; in mentions markets about −0.23 a day out.
E. Pooling twelve horizons from 1 to 90 days: in this simpler specification (no dollar-volume or staleness controls) the traders effect does not change with the horizon. The gaps between elections and crypto, geopolitics, politics, tech and economy markets widen further from the event. The mentions gap is about −0.3 throughout.

A. Contribution to Brier skill, one observation per market

	(1) Type only	(2) + traders	(3) + volume	(4) + age, staleness, year	(5) Log score
Crypto (vs. elections)	-0.103*(0.057)	-0.133**(0.052)	-0.136***(0.052)	-0.194***(0.055)	-0.161***(0.045)
Culture (vs. elections)	-0.068*(0.039)	0.003(0.040)	0.008(0.040)	0.003(0.041)	-0.009(0.032)
Economy/Fed (vs. elections)	-0.099***(0.034)	-0.078**(0.034)	-0.078**(0.034)	-0.133***(0.035)	-0.146***(0.033)
Esports (vs. elections)	-0.166**(0.079)	-0.069(0.082)	-0.067(0.079)	-0.154*(0.084)	-0.137**(0.067)
Geopolitics (vs. elections)	-0.140***(0.038)	-0.202***(0.039)	-0.200***(0.038)	-0.280***(0.040)	-0.231***(0.032)
Mentions/tweets (vs. elections)	-0.372***(0.045)	-0.247***(0.045)	-0.219***(0.044)	-0.296***(0.047)	-0.294***(0.041)
Other (vs. elections)	-0.136**(0.064)	-0.082(0.057)	-0.113**(0.057)	-0.205***(0.061)	-0.206***(0.052)
Politics (vs. elections)	-0.044(0.050)	-0.073(0.049)	-0.078(0.049)	-0.118**(0.049)	-0.116***(0.038)
Sports (vs. elections)	-0.065**(0.033)	0.006(0.035)	0.013(0.034)	-0.003(0.039)	-0.000(0.031)
Tech/Science (vs. elections)	-0.006(0.051)	0.023(0.049)	0.024(0.048)	-0.072(0.049)	-0.073*(0.040)
log(1 + wallets traded so far)		0.102***(0.009)	0.066***(0.016)	0.111***(0.019)	0.097***(0.015)
log(1 + $ traded so far)			0.052***(0.009)	0.055***(0.010)	0.055***(0.008)
log(lifetime volume)			-0.048***(0.008)	-0.069***(0.009)	-0.073***(0.007)
log(1 + market age, days)				-0.069***(0.014)	-0.037***(0.010)
log(1 + days since last trade)				-0.084***(0.013)	-0.052***(0.010)
Resolved early (by-date question)				0.074**(0.031)	0.066***(0.024)
Constant	0.619***(0.019)	0.169***(0.044)	0.359***(0.061)	1.043***(0.136)	0.924***(0.104)
Event-type effects	yes	yes	yes	yes	yes
Year effects				yes	yes
Markets	39,786	39,323	39,323	39,323	39,323
Event clusters	9,313	9,249	9,249	9,249	9,249
R²	0.002	0.010	0.012	0.017	0.021
Mean of dependent variable	0.550	0.548	0.548	0.548	0.531

One observation per market, 7 days before the event. Dependent variable: the market's contribution to Brier skill, [(base − outcome)² − (price − outcome)²] divided by its event type's mean (base − outcome)², where base is the type's share of Yes outcomes. Its mean within a type is that type's Brier skill, so coefficients read in Brier-skill units. Column 5 does the same with the log score. Standard errors clustered by event. *** p<0.01, ** p<0.05, * p<0.1.

	(1) Type only	(2) + traders	(3) + volume	(4) + age, staleness, year	(5) Log score
Crypto (vs. elections)	0.074(0.046)	0.059(0.043)	0.049(0.043)	-0.063(0.043)	-0.028(0.034)
Culture (vs. elections)	0.012(0.033)	0.064*(0.034)	0.064*(0.033)	0.026(0.034)	0.030(0.027)
Economy/Fed (vs. elections)	0.000(0.028)	0.004(0.028)	-0.000(0.028)	-0.061**(0.029)	-0.067***(0.026)
Esports (vs. elections)	-0.151***(0.047)	-0.057(0.049)	-0.057(0.048)	-0.174***(0.052)	-0.134***(0.041)
Geopolitics (vs. elections)	0.043(0.034)	-0.010(0.035)	-0.006(0.034)	-0.111***(0.035)	-0.070**(0.029)
Mentions/tweets (vs. elections)	-0.342***(0.025)	-0.231***(0.026)	-0.197***(0.025)	-0.304***(0.028)	-0.265***(0.024)
Other (vs. elections)	-0.062(0.056)	-0.001(0.049)	-0.028(0.049)	-0.128**(0.052)	-0.130***(0.048)
Politics (vs. elections)	0.024(0.040)	0.019(0.039)	0.014(0.039)	-0.062(0.039)	-0.042(0.030)
Sports (vs. elections)	-0.127***(0.023)	-0.040(0.025)	-0.032(0.025)	-0.106***(0.027)	-0.074***(0.021)
Tech/Science (vs. elections)	0.101**(0.042)	0.119***(0.041)	0.115***(0.040)	-0.001(0.042)	0.028(0.033)
log(1 + wallets traded so far)		0.091***(0.006)	0.043***(0.011)	0.078***(0.012)	0.055***(0.010)
log(1 + $ traded so far)			0.058***(0.007)	0.062***(0.007)	0.068***(0.006)
log(lifetime volume)			-0.040***(0.006)	-0.058***(0.007)	-0.067***(0.005)
log(1 + market age, days)				-0.056***(0.008)	-0.021***(0.006)
log(1 + days since last trade)				-0.095***(0.011)	-0.051***(0.008)
Resolved early (by-date question)				0.140***(0.020)	0.123***(0.015)
Constant	0.638***(0.018)	0.220***(0.034)	0.323***(0.045)	0.926***(0.105)	0.764***(0.087)
Event-type effects	yes	yes	yes	yes	yes
Year effects				yes	yes
Markets	61,518	61,431	61,431	61,431	61,431
Event clusters	15,492	15,475	15,475	15,475	15,475
R²	0.006	0.014	0.016	0.023	0.029
Mean of dependent variable	0.558	0.558	0.558	0.558	0.537

One observation per market, 1 day before the event. Dependent variable: the market's contribution to Brier skill, [(base − outcome)² − (price − outcome)²] divided by its event type's mean (base − outcome)², where base is the type's share of Yes outcomes. Its mean within a type is that type's Brier skill, so coefficients read in Brier-skill units. Column 5 does the same with the log score. Standard errors clustered by event. *** p<0.01, ** p<0.05, * p<0.1.

	(1) Type only	(2) + traders	(3) + volume	(4) + age, staleness, year	(5) Log score
Crypto (vs. elections)	-0.159**(0.076)	-0.226***(0.073)	-0.247***(0.074)	-0.300***(0.076)	-0.251***(0.060)
Culture (vs. elections)	-0.105**(0.051)	-0.038(0.055)	-0.026(0.055)	-0.054(0.059)	-0.056(0.045)
Economy/Fed (vs. elections)	-0.102(0.065)	-0.164**(0.067)	-0.162**(0.067)	-0.198***(0.068)	-0.182***(0.054)
Esports (vs. elections)	-0.166(0.130)	-0.026(0.133)	-0.014(0.129)	-0.097(0.133)	-0.095(0.107)
Geopolitics (vs. elections)	-0.277***(0.063)	-0.401***(0.062)	-0.413***(0.061)	-0.434***(0.064)	-0.357***(0.050)
Mentions/tweets (vs. elections)	-0.368***(0.097)	-0.213**(0.094)	-0.178*(0.094)	-0.177*(0.100)	-0.193**(0.095)
Other (vs. elections)	-0.294(0.194)	-0.191(0.179)	-0.215(0.175)	-0.239(0.180)	-0.261*(0.158)
Politics (vs. elections)	-0.085(0.094)	-0.188**(0.082)	-0.206**(0.083)	-0.232***(0.084)	-0.212***(0.063)
Sports (vs. elections)	-0.095*(0.050)	-0.046(0.052)	-0.034(0.051)	-0.118**(0.057)	-0.098**(0.046)
Tech/Science (vs. elections)	-0.068(0.087)	-0.082(0.087)	-0.085(0.086)	-0.142*(0.082)	-0.135**(0.064)
log(1 + wallets traded so far)		0.122***(0.014)	0.029(0.029)	0.029(0.035)	0.033(0.027)
log(1 + $ traded so far)			0.064***(0.014)	0.055***(0.014)	0.050***(0.012)
log(lifetime volume)			-0.016(0.016)	-0.033*(0.019)	-0.039***(0.015)
log(1 + market age, days)				0.014(0.019)	0.030**(0.015)
log(1 + days since last trade)				-0.089***(0.018)	-0.064***(0.014)
Resolved early (by-date question)				0.192***(0.047)	0.165***(0.036)
Constant	0.587***(0.023)	0.089(0.061)	0.100(0.100)	0.966***(0.265)	0.877***(0.200)
Event-type effects	yes	yes	yes	yes	yes
Year effects				yes	yes
Markets	21,415	20,354	20,354	20,354	20,354
Event clusters	4,941	4,823	4,823	4,823	4,823
R²	0.002	0.010	0.011	0.015	0.019
Mean of dependent variable	0.499	0.498	0.498	0.498	0.494

One observation per market, 30 days before the event. Dependent variable: the market's contribution to Brier skill, [(base − outcome)² − (price − outcome)²] divided by its event type's mean (base − outcome)², where base is the type's share of Yes outcomes. Its mean within a type is that type's Brier skill, so coefficients read in Brier-skill units. Column 5 does the same with the log score. Standard errors clustered by event. *** p<0.01, ** p<0.05, * p<0.1.

B. Calibration: outcome on price, 1 week before

	(1) Pooled	(2) × traders	(3) × thin	(4) × mentions	(5) All + type
Price	0.985***(0.006)	0.979***(0.006)	1.016***(0.006)	0.996***(0.006)	0.984***(0.006)
log wallets (centered)		-0.004***(0.001)			-0.006***(0.001)
Price × log wallets (centered)		0.039***(0.004)			0.037***(0.004)
Thin market (bottom wallet tercile)			0.009***(0.002)
Price × thin market			-0.107***(0.015)
Mentions market				-0.007(0.020)	-0.006(0.010)
Price × mentions market				-0.126***(0.041)	-0.080**(0.041)
Constant	0.001(0.001)	0.002*(0.001)	-0.001(0.001)	0.002(0.001)	0.001(0.002)
Event-type effects					yes
Markets	39,786	39,323	39,786	39,786	39,323
Event clusters	9,313	9,249	9,313	9,313	9,249
R²	0.553	0.558	0.555	0.554	0.559

One observation per market, 1 week before. Linear probability model of the outcome (1 = Yes) on the price. A calibrated market has constant 0 and price coefficient 1; a coefficient below 1 means prices are too extreme. Interactions show where prices are more or less informative. Standard errors clustered by event.

C. Information arriving in the last month

	(1) Type only	(2) + traders	(3) + controls
Crypto (vs. elections)	0.032***(0.007)	0.033***(0.007)	0.028***(0.008)
Culture (vs. elections)	0.004(0.003)	0.003(0.003)	0.000(0.004)
Economy/Fed (vs. elections)	0.011*(0.006)	0.014**(0.006)	0.010(0.007)
Esports (vs. elections)	0.018(0.011)	0.015(0.011)	0.008(0.012)
Geopolitics (vs. elections)	0.037***(0.007)	0.035***(0.006)	0.027***(0.006)
Mentions/tweets (vs. elections)	0.045***(0.017)	0.040**(0.017)	0.026(0.016)
Other (vs. elections)	0.056**(0.024)	0.046*(0.024)	0.036(0.023)
Politics (vs. elections)	0.011**(0.005)	0.012**(0.005)	0.011**(0.005)
Sports (vs. elections)	0.006**(0.003)	0.006**(0.003)	0.009**(0.004)
Tech/Science (vs. elections)	0.038***(0.010)	0.036***(0.010)	0.035***(0.009)
log(1 + wallets traded so far)		-0.005***(0.001)	-0.003(0.002)
log(1 + $ traded so far)			-0.001(0.001)
log(lifetime volume)			0.005***(0.001)
log(1 + market age, days)			-0.010***(0.001)
Resolved early (by-date question)			0.004(0.003)
Constant	0.021***(0.002)	0.041***(0.004)	0.007(0.012)
Event-type effects	yes	yes	yes
Year effects			yes
Markets	21,415	20,354	20,354
Event clusters	4,941	4,823	4,823
R²	0.007	0.009	0.019
Mean of dependent variable	0.030	0.030	0.030

One observation per market priced both 30 days and 1 day before. Dependent variable: fall in squared forecast error over the last month, (price₃₀ − outcome)² − (price₁ − outcome)². Positive means the price moved toward the outcome. Trading activity is measured 30 days before. Standard errors clustered by event.

E. How the effects change with time to the event (pooled horizons)

	(1) Day + type effects	(2) × log days	(3) + type × log days, controls
log(1 + wallets traded so far)	0.104***(0.007)	0.104***(0.008)	0.178***(0.011)
log wallets × log(days / 7)		0.001(0.004)	0.002(0.004)
log(lifetime volume)			-0.031***(0.007)
log(1 + market age, days)			-0.066***(0.008)
Crypto (vs. elections)	-0.099**(0.043)	-0.099**(0.043)	-0.150***(0.045)
Culture (vs. elections)	-0.001(0.036)	-0.001(0.036)	0.009(0.037)
Economy/Fed (vs. elections)	-0.087***(0.031)	-0.087***(0.031)	-0.116***(0.033)
Esports (vs. elections)	-0.090(0.065)	-0.091(0.065)	-0.107(0.072)
Geopolitics (vs. elections)	-0.190***(0.034)	-0.190***(0.034)	-0.251***(0.036)
Mentions/tweets (vs. elections)	-0.251***(0.028)	-0.252***(0.029)	-0.294***(0.040)
Other (vs. elections)	-0.065(0.052)	-0.065(0.052)	-0.138**(0.058)
Politics (vs. elections)	-0.096**(0.047)	-0.096**(0.047)	-0.116**(0.049)
Sports (vs. elections)	-0.029(0.032)	-0.029(0.032)	0.005(0.032)
Tech/Science (vs. elections)	0.017(0.043)	0.017(0.043)	-0.043(0.044)
Crypto × log(days / 7)			-0.084***(0.020)
Culture × log(days / 7)			-0.032**(0.014)
Economy/Fed × log(days / 7)			-0.056***(0.019)
Esports × log(days / 7)			-0.015(0.028)
Geopolitics × log(days / 7)			-0.087***(0.016)
Mentions/tweets × log(days / 7)			-0.021(0.021)
Other × log(days / 7)			-0.031(0.035)
Politics × log(days / 7)			-0.061***(0.019)
Sports × log(days / 7)			-0.008(0.011)
Tech/Science × log(days / 7)			-0.061***(0.021)
Event-type effects	yes	yes	yes
Year effects			yes
Markets	399,735	399,735	399,735
Event clusters	15,475	15,475	15,475
R²	0.010	0.010	0.013

One observation per market and horizon, at 1, 2, 3, 5, 7, 10, 14, 21, 30, 45, 60 and 90 days before the event. Dependent variable as in table A, with the base rate computed by type and horizon. Interactions use log(days / 7), so main effects are at one week, and a positive interaction means the effect is larger further from the event. Day effects included. Standard errors clustered by event.

D. Efficiency: does last week's price move predict the outcome?

	(1) Pooled	(2) × thin	(3) × mentions	(4) + type
Price	1.017***(0.006)	1.015***(0.006)	1.024***(0.006)	1.022***(0.006)
Price change over previous week	-0.151***(0.023)	-0.083***(0.029)	-0.159***(0.023)	-0.155***(0.023)
Thin market (bottom wallet tercile)		-0.009***(0.003)
Price change × thin market		-0.143***(0.045)
Mentions market			-0.099***(0.023)
Price change × mentions market			-0.020(0.131)
Constant	-0.004***(0.001)	-0.001(0.001)	-0.003***(0.001)	-0.004**(0.002)
Event-type effects				yes
Markets	31,500	31,500	31,500	31,500
Event clusters	7,288	7,288	7,288	7,288
R²	0.597	0.598	0.599	0.600

One observation per market priced 7 and 14 days before. If prices are efficient, the past week's change adds nothing once today's price is known (coefficient 0). Positive means prices underreact (moves continue); negative means they overreact (moves reverse). Standard errors clustered by event.

	(1) Pooled	(2) × thin	(3) × mentions	(4) + type
Price	1.016***(0.005)	1.014***(0.005)	1.020***(0.005)	1.017***(0.005)
Price change over previous week	-0.073***(0.014)	-0.034**(0.016)	-0.069***(0.014)	-0.075***(0.014)
Thin market (bottom wallet tercile)		-0.005**(0.002)
Price change × thin market		-0.111***(0.027)
Mentions market			-0.047**(0.020)
Price change × mentions market			-0.165**(0.069)
Constant	-0.004***(0.001)	-0.002*(0.001)	-0.004***(0.001)	-0.003**(0.001)
Event-type effects				yes
Markets	38,437	38,437	38,437	38,437
Event clusters	9,003	9,003	9,003	9,003
R²	0.672	0.672	0.672	0.673

One observation per market priced 1 and 8 days before. If prices are efficient, the past week's change adds nothing once today's price is known (coefficient 0). Positive means prices underreact (moves continue); negative means they overreact (moves reverse). Standard errors clustered by event.

Quoted prices against traded prices

Polymarket's price-history endpoint returns a quoted price that tracks the order book. In a market nobody is trading, the book can be a stray bid near 1¢ and an ask near 95–99¢, so the quote sits just under 50¢ regardless of the odds. Those quotes pulled the 45–50¢ bins far below the 45° line: a week out, quoted prices of 45–50¢ won only 32% of the time. With the last traded price, the same band wins 45%, and the dip disappears.

Quoted price, 1 week before41,566 markets

Last trade, 1 week before39,786 markets

Quoted price, 1 day before61,446 markets

Last trade, 1 day before61,518 markets

	Markets	Brier skill	Slope	Priced 45–50¢	Their win rate
1 week · Quoted price	41,566	0.530	1.00	1,240	0.32
1 week · Last trade	39,786	0.553	0.98	535	0.45
1 day · Quoted price	61,446	0.570	1.01	1,573	0.42
1 day · Last trade	61,518	0.565	0.99	1,220	0.47

Robustness

The main patterns survive each check. Weighting events equally, so a 20-candidate race counts once rather than 20 times, changes skill by up to 0.04. Dropping markets that resolved more than a day before their scheduled end, mostly "will X happen by date" questions that resolved Yes early, lowers skill by up to 0.05 and the slope to 0.89–0.96. By year of resolution, skill a week out is 0.55 in 2024 and 0.54–0.56 in 2025–26; a month out it fell from 0.56 in 2024 to 0.49 in 2026, as the platform grew into more and thinner markets.

Weighting each event equally (multi-candidate events split their weight)

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
3 months	10,059	2,053	0.12	0.4380.395–0.482	0.4110.372–0.449	0.9280.920–0.935	0.910.86–0.96	0.0150.010–0.021	0.0580.005–0.115
2 months	13,485	2,927	0.13	0.4500.416–0.488	0.4250.396–0.459	0.9360.929–0.943	0.930.89–0.97	0.0160.010–0.021	0.0500.010–0.088
1 month	21,415	4,941	0.15	0.4910.466–0.515	0.4640.441–0.486	0.9390.934–0.944	0.950.92–0.97	0.0070.003–0.011	0.0270.006–0.053
3 weeks	26,285	6,073	0.16	0.5130.487–0.535	0.4840.462–0.506	0.9400.935–0.944	0.980.96–1.00	0.0040.001–0.008	0.008-0.006–0.027
2 weeks	31,500	7,288	0.17	0.5320.514–0.551	0.4970.477–0.512	0.9410.937–0.946	1.000.98–1.01	-0.001-0.004–0.003	0.0240.008–0.043
1 week	39,786	9,313	0.18	0.5530.538–0.568	0.5120.497–0.528	0.9450.941–0.948	1.000.98–1.01	-0.005-0.008–-0.002	0.0200.007–0.034
3 days	51,220	12,020	0.20	0.5370.519–0.550	0.4910.473–0.504	0.9450.942–0.948	0.970.95–0.98	-0.009-0.013–-0.007	0.0320.018–0.045
2 days	55,415	13,323	0.21	0.5300.516–0.541	0.4850.470–0.497	0.9450.942–0.947	0.970.96–0.98	-0.009-0.013–-0.006	0.0280.017–0.038
1 day	61,518	15,492	0.23	0.5240.512–0.538	0.4820.470–0.495	0.9460.943–0.948	0.980.97–0.99	-0.010-0.013–-0.008	0.0130.006–0.020

Dropping markets that resolved more than a day before their scheduled end

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
3 months	5,223	1,598	0.14	0.4490.404–0.489	0.4410.402–0.475	0.9220.908–0.932	0.900.86–0.94	0.0050.000–0.010	0.019-0.016–0.063
2 months	7,280	2,285	0.14	0.4530.410–0.489	0.4450.406–0.478	0.9260.913–0.935	0.890.85–0.93	0.0060.003–0.010	0.028-0.004–0.062
1 month	12,596	3,892	0.15	0.4660.436–0.492	0.4560.432–0.479	0.9260.919–0.933	0.900.88–0.93	0.002-0.001–0.005	0.0310.011–0.056
3 weeks	15,471	4,761	0.16	0.4790.451–0.503	0.4690.445–0.490	0.9300.923–0.936	0.920.90–0.95	0.002-0.000–0.005	0.0270.008–0.049
2 weeks	18,510	5,633	0.17	0.4980.476–0.519	0.4840.467–0.505	0.9320.927–0.937	0.940.92–0.96	0.000-0.002–0.003	0.0260.010–0.039
1 week	23,087	7,065	0.19	0.5310.513–0.549	0.5100.493–0.526	0.9380.933–0.942	0.960.94–0.97	-0.000-0.003–0.001	0.0120.001–0.025
3 days	30,683	8,950	0.21	0.5250.513–0.540	0.5020.491–0.516	0.9350.932–0.939	0.960.95–0.97	-0.001-0.003–0.001	0.0220.011–0.034
2 days	33,496	9,659	0.22	0.5210.508–0.535	0.4990.487–0.512	0.9340.931–0.938	0.960.95–0.97	-0.002-0.004–-0.000	0.0210.010–0.030
1 day	37,274	10,643	0.23	0.5190.502–0.533	0.4960.479–0.509	0.9330.928–0.936	0.960.94–0.97	-0.003-0.005–-0.001	0.0220.011–0.032

By year of resolution, 1 week before

	Markets	Events	Base rate	Brier skill	Log skill	AUC	Slope	Gap <10¢	Gap >90¢
2023	201	180	0.27	0.6330.516–0.723	0.5960.486–0.683	0.9540.921–0.979	1.081.00–1.15	0.008-0.026–0.027	-0.028
2024	4,388	1,840	0.20	0.5510.516–0.594	0.5180.484–0.558	0.9390.930–0.949	0.980.95–1.01	0.002-0.003–0.007	0.0320.006–0.057
2025	14,168	3,548	0.17	0.5400.516–0.563	0.5280.506–0.551	0.9440.938–0.950	0.980.96–1.00	-0.001-0.003–0.002	0.005-0.008–0.023
2026	21,029	3,866	0.19	0.5600.539–0.576	0.5400.523–0.556	0.9460.942–0.950	0.990.97–1.01	-0.002-0.004–0.000	0.0110.000–0.025

How this was built

Sample. All closed Polymarket events from the public Gamma API (1.15 million events, through Oct 2026), keeping yes/no markets that resolved cleanly to 1/0 with at least $1,000 in volume, excluding auto-generated recurring series. 61,518 markets in 15,492 events had at least one trade in the 90 days before the event (resolving Oct 2023–Oct 2026).

Prices. For each market, the full trade history up to the event from the Data API (up to 10,000 trades paged back in time; for the 3% of markets with more, each older day is queried separately). The price on day d is the last trade at or before d days before the event, expressed as the price of the first listed outcome (usually "Yes"). Trades, distinct wallets and dollars traded are counted up to the same moment.

Event time. The earlier of the scheduled end date and the actual close. Polymarket's close time alone lags the real-world event by 12–20 hours at the median, because of the resolution process.

Uncertainty. 95% intervals come from 200 bootstrap draws that resample whole events (100 for the daily curves); regression standard errors are clustered by event.

Next steps. Separate the effect of traders from the selection of questions traders choose, for example with events that split into near-identical markets. Look at the price paths of mentions markets in their final hours. Model favourites' overpricing at long horizons, which may reflect the cost of capital locked up in a contract that pays no interest.

AUC: The chance that a market that resolved Yes was priced above one that resolved No. 0.5 is a coin flip, 1 is perfect ranking. Pure discrimination; ignores calibration.
Brier skill: The headline number. 1 minus the mean squared error of price as a forecast, relative to always forecasting the group's base rate. It equals (resolution − miscalibration) / uncertainty, so it reads like an R²: the share of outcome uncertainty the price resolves, net of bias. 0 means no better than the base rate.
Log skill: The same idea with the log score (a McFadden pseudo-R²), prices clipped to 1–99¢. It punishes confident prices that turn out wrong much harder than Brier skill does.
Slope: Regression of the outcome on price. 1 means calibrated; below 1 means prices are too extreme, above 1 means too timid.
Gap <10¢ / >90¢: Mean price minus win rate among contracts priced under 10¢ (or over 90¢). Positive under 10¢ means longshots are overpriced; negative over 90¢ means favourites are underpriced.