Polymarket and Kalshi Price Discovery: A Three-Day Pilot

Correction and Update — 19 October 2026

We are withdrawing this pilot post's directional conclusion ("Polymarket leads, Kalshi follows." — 63.6% of follow-events across 3 days of June 2026 MLB). A much larger August–October 2026 study across five sports and 581 games finds the opposite direction at the 32-second horizon (Kalshi leads, −0.304), and shows on second-resolution websocket data that the two venues move within about 1 second of each other, so neither "leads" in a tradeable way.

• What we said (24 June 2026): On 26–29 settled MLB games over 3 days (June 21–23, 2026; ~57K snapshots joined at a ~31-second median cadence), Polymarket led 63.6% of follow-events, had a 3.8× Granger cross-coefficient asymmetry (0.46 vs 0.12), and led in 25 of 26 games.
• What later data shows (18 October 2026): Across 581 matched games over 12 UTC days (2026-08-29 to 2026-10-16; CFB 220, MLB 112, NBA 62, NFL 84, NHL 103; 194,408 32-second intervals pairing Polymarket depth snapshots against exchange-timestamped Kalshi websocket updates), the cross-correlation lead-lag asymmetry is −0.304 (95% CI [−0.342, −0.269]), with Kalshi leading in every sport (CFB −0.355, NFL −0.366, MLB −0.267, NBA −0.209, NHL −0.192). Polymarket closes 32% of an instantaneous cross-venue gap over the next 32s sample while Kalshi closes 4%. Against an exchange-timestamped NBA websocket (1,724 snapshots), Polymarket's REST depth timestamp matches the live websocket state within 0–2 seconds (78.9% exact BBO match at +2s, mean |mid diff| = 0.16c), ruling out a 30-second Polymarket clock bug. On 1-second dual-websocket NBA data (18 series, 28.5 hours), both venues reprice within ~1 second of each other.
• What we now believe: A ~30-second polled grid with symmetric nearest-as-of matching (±30s) cannot reliably order venues that move within ~1 second; on 1-second data neither venue leads in a tradeable way, and at the 5–30 second horizon in slower sports books in our August–October 2026 data, Kalshi leads and Polymarket follows. The section "Why you can't trade this" below is affirmed by both studies: taker strategies chasing cross-venue lags lose after spread and fees in both directions.
• What is unknown: How much of the sign reversal between June–July MLB and August–October MLB comes from the earlier symmetric nearest-as-of join (±30s) versus a possible market-structure change since July — Polymarket sports markets now hold taker orders 1–3 s (secondsDelay; whether this was in force in June is unverified) and charge a taker fee of about 1–2.5% of notional at mid-range prices (0.75–1.25c per share at 50c, from rate parameter 0.03–0.05 in rate * p * (1 - p)) — as well as sub-second ordering outside NBA where dual websocket capture is not yet available.

See the full comparison table in the updated 16-day lead-lag study and our CLV retraction. The original June 2026 pilot text is preserved intact below.

When the same MLB game trades on both Polymarket and Kalshi, the two prices move almost in lockstep — they correlate 0.996. But "almost" is where the interesting question lives: when they do move, which venue moves first? Which book is leading price discovery, and which is following?

The original June analysis reported a Polymarket-first pattern in the aligned sampled prices. The tests below use the same underlying observations; they are not independent experiments. This update preserves their historical values while separating sampled association from exchange-time causal price discovery. The original payload and clock/alignment provenance were not independently rerun for this update.

The result

We took a matched book of 26–29 settled MLB games over 3 contiguous days (June 21–23, 2026; ~57K snapshots, ~31-second median sampling) and looked at "follow events": moments where one venue's mid moves while the other is momentarily flat. The question is simply which venue tends to be the one that moved.

• Polymarket leads 63.6% of follow-events (p < 1e-6).
• The direction replicates every single day: 0.657 / 0.667 / 0.635 across the three.
• Three methods on the same observations agree:
• Granger cross-coefficient asymmetry of 3.8× (0.46 vs 0.12)
• lagged cross-correlation 0.296 vs 0.072
• per-game clustering: 25 of 26 games show Polymarket leading (Wilcoxon p < 1e-4) — a cross-game pattern in that reported sample, not proof that timing and alignment artifacts have been eliminated.

Agreement across methods is useful, but the methods share the same prices, clocks and matching choices. Three days do not establish a stable venue lead. We later re-ran the question on a sixteen-day window — the follow-up lead-lag study found the same direction on all sixteen days.

What about which venue is right?

A natural follow-up: if Polymarket moves first, is Kalshi the more accurate of the two (since it has the extra beat to incorporate information)? The data hints at this — Kalshi's level is fractionally closer to the final settlement — but the honest reading is that this half is weak and we won't headline it:

• the per-game-team accuracy gap is only 0.125¢ (Brier difference 0.0013),
• the paired t-test is p = 0.056 — not significant at the conventional 0.05 bar,
• and it does not replicate: drop June 23 and the tally flips to 12 Kalshi-closer / 13 Polymarket-closer. The entire "accuracy edge" is essentially carried by a single day.

The reported lag pattern and accuracy comparison are separate observations. Neither establishes a persistent executable trading advantage.

Why you can't trade this

This is the part a less honest write-up would leave out. A 1¢ lead is not a 1¢ profit. The original comparison used a 3–5c round-trip cost assumption. That is not a universal venue fee or a fill-by-fill cost measurement. A quote relationship alone does not establish an executable edge.

The original article also cited a directional-bot snapshot of −6.7c CLV, −5.7% ROI and +2.2c slippage. Those are historical reported values, not current account statistics, and a separate bot is not a control for this lead-lag inference. See the dated CLV correction and current descriptive scorecard for the measurement context.

What we're not claiming

• N = 26–29 settled games over 3 contiguous regular-season MLB days, one capture window (with a known Apr 22–Jun 14 gap).
• The ~31-second median sampling floor means this is a 30–90-second-horizon statement, not a microsecond price-discovery claim. The grid cannot resolve an exchange-time lead shorter than its observation spacing. Fresh clocks and matching rules must be checked before interpreting the direction.
• We track one team/side per game — we see half of each market.
• Granger t-statistics are inflated by pooled autocorrelation, so the load-bearing evidence is the 3.8× coefficient asymmetry plus the per-day and per-game replication, not raw t-values.

See for yourself

The free, score-synced matched-book sample — both venues on one timeline, winner labeled — is on Hugging Face and Zenodo (DOI 10.5281/zenodo.20816908). It lets you inspect a single-game matched schema. Reproducing the multi-game p-values above requires the exact original window, code and alignment inputs. The released matching schema permits nearest-time matches, including future observations; use a backward-only rule for a causal research design.

For the capture methodology, see how to get historical Polymarket order-book data, and if you're weighing all the sources — us included — start with our comparison of where to get historical Polymarket and Kalshi order-book data. To reproduce this cross-venue result, inspect the Polymarket–Kalshi MLB matched book.