Polymarket perps lead Hyperliquid by 250–500 ms on source clocks

43 of 44 same-symbol pairs peak with Polymarket first across 42 days of perps quotes, and Polymarket's returns forecast Hyperliquid's next half-second far better than the reverse.

We compared Polymarket's perpetual quotes with Hyperliquid's for 83 Polymarket instruments over 42 days, August 1 to September 11, 2026: crypto, US equities, commodities and indices. On each venue's own source timestamps, Polymarket moves first. The lead shows up in lag correlations and in out-of-sample forecasts: Polymarket's last few seconds of returns help predict Hyperliquid's next half-second far more often than the reverse.

Source timestamps, your latency. Every figure here uses the venues' own clocks: Polymarket's websocket frame timestamp and Hyperliquid's block time. That keeps our network path out of the measurement. It also means the lead is not something you get for free: what a trading bot actually sees depends on where it runs and how fast it receives both feeds and reaches the order book. Anyone trying to use this needs to measure and optimize their own latency. The cost section below shows how quickly the illustrated edge shrinks with delay.

Where the lead appears

The six examples were fixed before fitting. Across the full same-symbol universe, 1 pair peaks at zero lag; 38 pairs peak at a 250 ms Polymarket lead; 5 pairs peak at a 500 ms Polymarket lead. The training and evaluation peak bins agree for 41 of 44 pairs. Peak correlation lag is not a per-update reaction time. The single zero-lag pair is STRC. The full lag table includes return-clipping and quote-age sensitivity checks.

How much Polymarket adds to a Hyperliquid forecast

Each model predicts Hyperliquid's next 500 ms. The baseline uses only Hyperliquid's own trailing returns; the full model adds Polymarket's. Both are fitted on August 1–28 and evaluated, with frozen coefficients, on August 29–September 11.

R² measures squared-error skill against predicting no price change, not directional accuracy. On its own history, Hyperliquid's next half-second is close to unpredictable; adding Polymarket's returns lifts BTC and ETH above 30%. The last column is a cross-check: the same model refit on Hyperliquid quotes we captured directly from its websocket, instead of books rebuilt from node data. The two sources agree. Refits and clipped-return checks are in forecast_validation.csv.

The forecast test and quality gate

The baseline uses the target venue's trailing 0.5-, 1- and 5-second returns. The expanded model adds the other venue's returns over the same windows. Coefficients stay frozen after training; targets do not overlap within a specification. A strict pass requires improvement over both own-history and no-price-change baselines, positive day-level confidence limits, and Benjamini–Yekutieli q < 0.05 against both across 452 hypotheses. The full forecast table includes failures and effect sizes.

Each Hyperliquid coin-day must pass a quality gate: at least 100 checkpoints where the book's best bid and ask are compared with captured L2 snapshots, with at least 95% exact agreement on each price. 1,462 of 1,585 checked coin-days pass; the checkpoint table lists every one. This is exploratory inference on a six-week sample.

Spreads, fees and delay

A forecast can be informative while a trade loses money. This illustration acts when the frozen one-second forecast exceeds twice its training standard deviation, crosses the target spread on entry and exit, and holds for one second after an assumed source-time entry delay. Figures are basis points after spreads, before fees.

Every 250 ms of delay eats into the result, and a 2 bps round-trip fee erases the average ETH and SOL figure even at zero delay. These are quote-based illustrations, not fills or a profitable strategy. Whether the lead is usable comes down to your own feed and order latency, measured against one local clock.

What the clocks and sample establish

The two venues define time differently: Polymarket stamps the websocket frame, Hyperliquid stamps the block. Treat the 250–500 ms figure as a source-clock measurement of which market moves first, not as a guaranteed window. Book receive timestamps are not used.

Polymarket lists Hyperliquid among possible index-price inputs, with feed selection varying by session. This is context, not a causal explanation demonstrated by the study. We use book midquotes, not index prices, and do not know historical per-instrument feed weights.

The screen contains 83 PM instruments and 139 candidate comparisons. 49 pairs on 48 PM instruments meet at least one forecast specification; aliases and related share classes stay separate. Same symbols do not guarantee identical contract specifications. The coverage table reports available dates and quality counts. Fresh, positive, uncrossed quotes are required; stale windows and UTC boundaries are censored. Evaluation therefore conditions on observed coverage.

Six weeks cover a short regime. Resampling days and adjusting for multiple comparisons quantify uncertainty within this sample; they cannot establish persistence afterward. The reproduction includes finer timing grids, session and volatility splits, and all tested horizons, including specifications that did not meet the coverage requirement.

Data and reproduction

Download the code and results and cost scenarios. The pinned Docker workflow records input identities, software versions and checksums. Reproducing the calculations requires the historical data; raw market data and credentials are not included in the download.

TickFoundry collects and sells the datasets used in this research.