Skip to content

Latest commit

 

History

17 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

0xinsider research

The SQL, the raw query output and the bootstrap scripts behind the studies published at 0xinsider.com/research.

Every figure on those pages comes from a run committed here. If a number on the site does not match a number in this repository, the repository is right and the page has a bug worth an issue.

The studies

grade-vs-price — do wallet grades predict outcomes?

67,531 Polymarket buys of $10,000 or more between 2026-06-01 and 2026-09-11, each scored against the settled outcome using the grade the wallet already held on the day it traded.

Cohort Trades Markets Won Price paid Edge 95% CI
S/A/B 24,610 5,946 66.5% 65.0c +1.57 pts [+0.32, +2.82]
C 9,223 1,986 61.8% 61.3c +0.53 pts [-1.93, +2.81]
D/F 17,106 4,476 56.6% 58.7c -2.06 pts [-3.61, -0.45]
No grade 16,592 5,250 59.2% 59.3c -0.13 pts [-1.81, +1.57]

Published: https://0xinsider.com/research/do-wallet-grades-predict-outcomes

polymarket-sports-markets — fifteen studies on sports

411,770 sports buys of $1,000 or more between 2026-04-02 and 2026-09-11, $8.21B.

The later sports studies (the wallet census, fading the crowd, over/under totals, soccer draws, cashing out, both teams to score, the first set in tennis, point spreads, home advantage, NRFI, football underdogs and esports first-map winners) are listed with their files in the directory's README.

three-venue-prices — the same games on Polymarket, Kalshi and DraftKings

81 games priced on all three venues in one eight-second capture on 2026-09-18. Polymarket and Kalshi showed the identical midpoint on 57 and never differed by more than 2 points. DraftKings' moneyline with the margin removed sat a median of 0.99 points from Polymarket, and its overround was a median of 4.25 points against 3.16 cents for both sides on Polymarket with the taker fee. Public sources only: python3 analysis.py capture.json reproduces every figure with no database.

All 81 have since been played and are scored against their results: Brier 0.1502 on Polymarket, 0.1504 on Kalshi and 0.1510 on DraftKings with its margin out, against 0.2414 for the sample's own home rate. All three beat a no-information price by a wide margin and none is separable from the others -- every pairwise interval spans zero on 81 games. python3 scoring.py --games games.csv --results results.csv reproduces that too.

Published: https://0xinsider.com/research/use-0xinsider-for-kalshi-draftkings

Reproducing

Each study directory holds its SQL, the raw psql output of the run the figures come from, and a market-clustered bootstrap.

cd grade-vs-price
psql "$DATABASE_URL" -X -f queries.sql          # the printed tables
python3 bootstrap.py                            # the confidence intervals

grade-vs-price/market_edge.csv is committed, so the bootstrap runs without database access and reproduces the published intervals exactly:

cohort      trades  markets  edge_pts   95% CI
S/A/B        24610     5946      1.57   [+0.32, +2.82]
C             9223     1986      0.53   [-1.93, +2.81]
D/F          17106     4476     -2.06   [-3.61, -0.45]
no grade     16592     5250     -0.13   [-1.81, +1.57]

The sports bootstrap takes the market-level CSVs the \copy lines in each .sql file write, so it needs a database. bootstrap-output.txt is the run the published figures come from.

You cannot run the SQL against our database. It reads internal tables through a read-only role. The SQL is published so the method can be read and argued with, and so anyone holding comparable Polymarket data can run the same test.

What the numbers mean

Edge is the realized win rate minus the average price paid, in percentage points. Buy at 65c and win 66.5% of the time and you are 1.5 points better than the market that sold to you. Win rate alone measures which prices someone likes, not whether they are right.

Confidence intervals resample markets, not trades. Four wallets buying the same side of the same market are one observation, not four, and treating them as four is how a result gets manufactured. 5,000 draws, seed 20260912.

Grades are point-in-time. Each trade takes the most recent grade dated on or before the day of the trade. A grade computed after a market settled never touches the trade it would have predicted.

Things that will bite you

These are in the per-study READMEs too, because each one already produced a wrong answer once.

  • The sample grows. Every trade sits on a market that has settled, so the universe widens as open markets resolve. Three runs of the unchanged grade query ninety minutes apart returned 67,516, 67,517 and 67,531 trades. Bounding on resolved_at does not freeze it: rows gain an outcome after the fact carrying a timestamp that predates the bound. Pin a run timestamp and say so. The conclusions held across all three runs.
  • Grade coverage widened in June 2026, from a few hundred wallets scored per day in May to about 18,500 in June. A window reaching further back reads "no grade" as a fact about the wallet when it is a fact about us. An earlier pass made that mistake and produced a meaningless -11.11 for March.
  • Sports data before 2026-04-02 is unusable for this. The outcome index on large buys was a defaulted 0 for a large share of rows: 80c+ buys tagged outcome 0 "won" 53 to 56% in February and March while outcome 1 won 84 to 89%; from April 2 both are about 88%. An early pass read a 25-point reverse favorite-longshot bias that was this defect.
  • One market can carry a bucket. The 10-20c sports bucket shows +27.6% on the dollar, and $41.31M of its $25.89M total is one market, Spain to win the World Cup at 15.2c. The market-clustered interval catches it: [-5.64, +19.10].

Licence

CC BY 4.0. Use it, quote it, check it, argue with it. A link back to the study page is enough attribution.

Corrections are welcome as issues. Two number defects on the published pages were caught in one day by checking prose against query output rather than by proofreading, which is the check that works.

About

SQL, raw query output and market-clustered bootstraps behind the Polymarket studies published at 0xinsider.com/research

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages