Source correction — October 8, 2026: The earlier article gave venue-wide calibration, price-adjustment and liquidity rankings without a linked reproducible artifact for those comparisons. Those claims are withdrawn. The replacement below reports specific ZenHodl releases from the public catalog captured October 8, not a census of prediction markets or evidence of trading profit. The September retraction is retained verbatim below as a historical notice; its ledger totals are not October totals.
Correction (2026-09-25): Two claims below have since been contradicted by later, larger reviews and should not be read as current. "Our models detect approximately 5-15 actionable mispricings per game night" (Finding 2) and the "showed promise" win rates for hold-to-settlement moneyline trading (67% MLB / 64% NHL / 60% NBA) reflect this post's original 1,100-trade sample from March-April 2026. A September 2026 review of 2,584 live fills found no sport clears a positive-expectation edge floor in the 0-30¢ range, and the bought-side calibration gap (realized win rate minus market-implied probability) is negative in every sport, roughly −7pp to −31pp (internal fill study, September 2026). The canonical live ledger since reconciled shows a 45.9% win rate and a −$217.75 cumulative result over 2,214 live trades (detail at the infrastructure post's correction), not the sport-specific win rates quoted below. Treat "actionable mispricings" and "showed promise" as historical, not as a current claim that a tradable edge exists.
What our released files actually contain
The machine-readable catalog identifies the artifact version, row units, observation clocks, sample, schema and known limits for each product. Four releases illustrate why a generic “historical market data” label is inadequate.
| Release | Catalog count | Frozen version |
|---|---|---|
| Polymarket sports books | 1,628,787,386 quote + depth rows | through-2026-10-03-reviewed-20261004 |
| Kalshi depth and prints | 851,964,251 depth + print rows | through-2026-10-02-reviewed-20261003 |
| Matched MLB quotes | 1,394,808 matched rows | through-2026-10-02-reviewed-20261003 |
| Combined combo release | 16,561,512 rows across distinct tables | through-2026-10-04-reviewed-20261005 |
These counts use different row units and overlapping product populations. Do not add them together as unique observations, markets, trades or full-exchange volume.
The Polymarket sports release contains 1,041,590,302 quote rows and 587,197,084 depth rows. The quoted observation window is March 28–October 3, 2026; its L2 depth series starts June 16. It records capped ladders of up to 20 levels per side, not every book update or a trade-print tape. A recorder cycle of about 30 seconds is not a guarantee that each source quote is 30 seconds old.
The Kalshi tape contains 821,979,759 depth rows and 29,984,492 captured unique trade prints. Its book bounds are June 21 at 17:24:31.656958 UTC through October 2 at 23:59:16.002162 UTC. Capture selected markets, book polling and print retrieval have different populations and clocks. The catalog reports a global depth-cycle median of 30.5 seconds before July 12 and 43.0 seconds afterward; neither establishes continuous per-ticker coverage. The Kalshi schema documents the lost historical fractional-depth precision and separate exchange/retrieval timestamps.
A matched file is not simultaneous execution
The MLB release has 1,394,808 matched rows across 1,005 dated games: 965 settled and 40 unresolved in that artifact. Its last book observation is September 30, even though the artifact includes files through October 2. A nearest-time join can pair observations within ±30 seconds, including a future observation relative to the other venue. That is useful for quote comparisons; it is not proof of which venue moved first or that two legs could fill together. The matched-book schema identifies the independently observed side and derived complements.
Combo records need their own clocks
The combined release includes 1,323,840 Polymarket fill rows and 1,676,300 captured Kalshi combo prints, plus definition, leg and outcome tables. Its 16,561,512 total rows are not 16,561,512 economic trades. Polymarket fill block times, Kalshi exchange times, recorder retrieval and artifact inclusion dates are separate.
Polymarket combo book history is not included. Kalshi has one registration-time top-of-book snapshot per combo, not a repeated book series. The released coverage estimate is about 2% of Kalshi combo trades over August 6–October 4, with a reported 95% interval of 1.4–2.7%; it is a captured sample. See the combo schema, including block-time estimation and gap qualifications.
What this evidence supports
These releases support schema inspection, observation-gap analysis and scoped historical quote or print research. They do not establish full-season coverage, actual fill quality or a profitable strategy. The CLV retraction explains why a near-resolution close can encode the eventual outcome; calibration and a model-price gap alone cannot substitute for an execution study.
Start with free sample rows, compare their clocks and fields with your research question, and then inspect the specific release’s coverage notes. The tryout walkthrough demonstrates one dated ZIP rather than assuming every archive behaves the same way.