Name
Methodology and data
How hyperliquid.build collects order-book snapshots, defines sessions, computes spread, depth and slippage, and what the numbers cannot tell you.
Source
Our own collector polls the public Hyperliquid /info API for every market on the xyz HIP-3 deployer and stores the raw responses. The Liquidity Lab is built from:
l2Book: 20 price levels per side, one snapshot per market roughly every six minutes, continuously since 2026-06-16 (106 days at the last rebuild).metaAndAssetCtxs: mark price, oracle price, 24-hour notional volume, open interest and current funding, sampled about once a minute; we keep the last sample of each UTC day.fundingHistory: hourly settlement rates for the last 30 days.meta:szDecimals,maxLeverage,growthMode,deployerFeeScale,isDelistedfor every asset.
The site is rebuilt once a day. Numbers on every page are also available as static JSON under /data/xyz/; see the data terms below.
Sessions
HIP-3 markets trade around the clock. We convert each snapshot's exchange timestamp to New York time and assign one of four buckets:
| Session | Definition |
|---|---|
| US regular hours (RTH) | Monday–Friday 9:30–16:00, except NYSE holidays; on half days, 9:30–13:00 |
| US extended hours | 4:00–9:30 and 16:00–20:00 on trading days |
| Overnight | 20:00–4:00 between consecutive trading days |
| Weekend / holiday | Friday 20:00 to Monday 4:00, plus NYSE full-day holidays |
The 2026 holiday calendar used: 1 Jan, 19 Jan, 16 Feb, 3 Apr, 25 May, 19 Jun, 3 Jul, 7 Sep, 26 Nov, 25 Dec; half days 27 Nov and 24 Dec. The buckets describe US equity hours regardless of the asset, so for commodity, FX and non-US names they are a time-of-week label rather than a statement about the underlying market's hours.
Metrics
For every snapshot with a valid best bid and ask, with mid = (bid + ask) / 2:
- Spread:
(ask − bid) / midin basis points. - Top of book: best bid size plus best ask size, in USD.
- Depth within 10 / 25 / 50 bps: the notional resting at prices within that distance of mid, summed over both sides.
- Slippage for a given order size ($1k, $5k, $10k, $25k, $50k, $100k, $250k): we walk the visible levels from the best price outward until the notional is filled, compute the average fill price, and express it against mid in basis points. We do this for a buy (walking the asks) and a sell (walking the bids) and report the mean of the two. If the 20 visible levels cannot fill the size, the snapshot counts as not filled for that size; the fill rate is the share of snapshots in the session that could fill it.
- Per session: median and 90th percentile of the above over all snapshots in the session; medians for depth and top of book.
- Hour of week: median spread and median depth within 25 bps for each of the 168 hour-of-week cells in New York time.
- Weekly trend: median spread, depth within 25 bps and $25k slippage across all sessions per calendar week.
- Funding APR: mean hourly settlement rate over the last 30 days × 24 × 365, in percent; positive means longs pay shorts.
- Taker fee:
4.5 bps × multiplier × (0.1 if growth mode), where the multiplier is1 + deployerFeeScalefor scales up to 1 and2 × deployerFeeScaleabove. Maker fees, fee tiers and rebates are not modelled. - Average 24h volume, 30 days: mean of the last 30 daily samples of
dayNtlVlm. - Open interest:
openInterest × markPxat the last sample. Hyperliquid's open interest is the sum of longs and shorts.
The cost calculator on each market page combines these as 2 × slippage + 2 × taker fee per round trip, times round trips per day, plus daily funding for a position held all day. Slippage between the tested sizes is interpolated on a log scale.
Pre-market board
The pre-market board asks one question per NYSE trading day: while Wall Street was closed, which way did the on-chain price of each US stock and ETF perp on xyz move, and did the real opening gap go the same way? Only markets we classify as US equities or US ETFs are included. Pre-IPO and non-US names are excluded because they have no US open to compare against; indices, commodities and FX are out of scope for now.
All times are New York time (daylight saving handled). A trading day is a weekday that is not an NYSE holiday. The previous session closes at 16:00, or 13:00 on a half day; the next one opens at 09:30.
| Term | Plain-English definition |
|---|---|
| Window | From the previous trading day's close to 09:30 on the trading day. Called overnight when the two days are consecutive calendar days, weekend for Friday to Monday, and holiday for any other gap, including a weekend that runs into a holiday. |
| Reference close | The deployer's oracle price (its reference for the underlying stock) at the previous close: the last sample within a minute either side of the close, otherwise the last sample up to five minutes after it. |
| Real open | The oracle price at 09:30: the first sample between 09:30:00 and 09:31:59, otherwise the first sample up to 09:35. Empty until the market has opened. |
| Chain close | The close of the perp's 1-minute candle for the last minute of the previous session (15:59, or 12:59 on a half day), otherwise the last candle up to five minutes after the close. |
| Chain at 04:00 | The close of the 03:59 candle on the trading day, otherwise the last candle before 04:00 that is still after the previous close. After a weekend or holiday this is the trading day's 03:59, so it comes after the whole closed stretch and before the US pre-market opens. |
| Chain at 09:00 | The close of the 08:59 candle on the trading day, the last full minute before 09:00, otherwise the last candle before 09:00. |
| Chain at the open | The close of the 09:29 candle, otherwise the last candle before 09:30. |
| Early call (04:00 call) | Chain at 04:00 against chain close, in percent. US pre-market trading starts at 04:00, so this is the move made while no US venue was open. |
| Chain move (09:00 call) | Chain at 09:00 against chain close, in percent. This is "the chain's call", published before the open. By 09:00 the US pre-market has been trading for five hours and the perp tracks it, so much of this move repeats prices already set on US venues. |
| Chain now (so far) | The last candle close at or before the board's as-of time, against chain close. Before 09:00 the live board shows this in place of the 09:00 call. |
| High and low | The highest 1-minute high and lowest 1-minute low while the market was closed, against the chain close. |
| Path | The candle close every 15 minutes from the previous close to 09:30, against the chain close; gaps where no candle printed. |
| Real gap | Real open against reference close, in percent. |
| Agree | The chain move and the real gap have the same sign, counted only when both are at least 0.10% in size. Smaller moves fall in the dead band and are not scored. A second figure ("no band") counts every night with both values. The early call is scored the same way (early agree), separately. |
| Overnight volume | The sum of each 1-minute candle's base volume times its close across the window; an approximate notional. |
| Data OK | False when the reference close, the chain close or the latest chain price (chain now) is missing. Such rows stay on the board with blanks and are left out of every rate. A call that is still missing (for example the 09:00 call before 09:00) is simply not scored. |
The scorecard aggregates every completed night since 2026-06-17, separately for the 04:00 and the 09:00 call: the share of calls that agreed, the mean absolute difference between the chain move and the real gap (in percentage points), the medians of both moves, and beta, the least-squares slope of the real gap on the chain move with no intercept (1 means the gap matched the chain move on average; below 1 means the open tended to move less than the chain had). Beta is reported only from 15 observations. Breakdowns are by window type, by the size of each call's own move, and by symbol; the calibration charts plot every completed symbol-night for each call, up to the latest 6,000. The 09:00 call agrees with the open far more often than the 04:00 call, which is expected: it has had five hours of US pre-market to follow. The 04:00 call is the part only a 24/7 venue can provide.
Rules we hold the page to:
- Every rate shows its N, and no rate is shown below n = 15. Small samples are listed with their count and no percentage.
- The 0.10% dead band is stated next to every agreement rate.
- The scorecard is a description, not a forecast. It says how often the chain's direction matched the open on past nights. It does not say what the next open will do.
- 1-minute candles can be sparse overnight. A quiet market may print no candle for long stretches, which leaves holes in its path and can push the chain close or the 09:00 price onto an earlier candle.
- The oracle is sampled about once a minute. The reference close and real open approximate the official closing and opening prints; they are not the exchange's auction prices.
- Pre-IPO and non-US names are excluded because they have no US open.
The board is published twice: the nightly rebuild at 00:40 UTC recomputes everything and fills in the real open for the night before, and a second run shortly after 09:00 New York time publishes the board as of 09:00, before the open. Each trading day is archived at /premarket/<date>/ and as JSON at /data/xyz/premarket/<date>.json.
What the numbers cannot tell you
- Six-minute snapshots miss what happens between them. Bursts of liquidations, news spikes and the seconds around the US open and close are under-sampled. Medians are robust to this; p90 values are a floor, not a ceiling, for bad moments.
- Twenty visible levels are not the whole book. Large resting orders often sit outside them, and most walls last minutes. Slippage here is what a market order would pay against the visible book at that instant, nothing more.
- No hidden liquidity, no impact after the fill. Market makers refill; large orders move the price for longer than one snapshot. Real costs for repeated large orders will exceed these figures.
- Fees are the base formula only. Your tier, rebates and referral settings change the number.
- Funding is backward-looking. A 30-day mean is a description, not a forecast, and sign flips are common.
- Known collection gaps: the funding stream has gaps between 2026-08-28 and 2026-09-11; order-book polling lost about a minute on 2026-09-28 during a rate-limit event on our side. Days missing entirely from
l2Bookwould show as missing weeks in the trend charts; none were missing at the last rebuild.
Data terms
The JSON under /data/ is free to use in your own bots, dashboards, research and articles, on two conditions:
- Attribution: cite hyperliquid.build with a link wherever the numbers appear.
- No bulk redistribution: do not mirror, resell or republish the dataset as a whole, and do not serve our files to others as your own API. Link to us instead; the files are rebuilt daily and a copy goes stale.
Requests to /data/ are rate-limited per IP at the edge. If you need the full history or a higher limit, ask; that is what the builder account on the roadmap is for.
Independence
hyperliquid.build is an independent project. It is not affiliated with, endorsed by, or operated by Hyperliquid Labs, the Hyper Foundation, or any HIP-3 deployer. It sells nothing, holds no positions in the markets it measures at the time of writing, and does not offer investment advice. Questions and corrections: open an issue once the repository is public, or write to the address in the site footer when one is listed.