Every few weeks someone posts that Hyperliquid quotes lag the big centralized books, and that the gap is there to be taken. The claim is testable and almost never tested. So I tested the first half of it, the part that decides whether the second half is even measurable: how often does each venue print a new top of book, and how far apart are the four mids?

The probe is 300 lines of standard-library Python with no key and no account. I ran it from a home connection in Seoul, captured six seconds of BTC, and printed the table below. Everything that follows comes from that capture, which ships with the repository so you can replay it without opening a socket.

The update clock

venue status connect first quote quotes upd/s median gap spread med spread min bid ask --------------- ------ ------- ----------- ------ ----- ---------- ---------- ---------- -------- -------- hyperliquid ok 224 ms 368 ms 65 11.0 75 ms 0.16 0.16 62,507.0 62,508.0 binance_futures ok 331 ms 365 ms 2,247 376.5 <1 ms 0.02 0.02 62,506.0 62,506.1 okx_swap ok 510 ms 890 ms 352 62.4 11 ms 0.02 0.02 62,509.4 62,509.5 bybit_linear ok 520 ms 661 ms 276 47.0 10 ms 0.02 0.02 62,502.1 62,502.2 

Read the fourth column against the fifth. Hyperliquid delivered 11 updates per second on BTC while Binance delivered 376. The target venue's own quote clock is a floor on any claim that it lags someone: if a book prints a new touch every 75 ms at the median, a 250 ms window difference is measuring that venue's tick granularity, not a delay.
The four mids sat 1.17 bps apart at the end of the window. A round trip on those venues costs several times that. That single line is most of the practical story, and the rest of this post is the arithmetic behind it.

Why an 88 percent hit rate still loses money

The second half of the claim needs a definition before it needs data. For a window w and an anchor t on Hyperliquid's quote stream:

  • r_L is the leader's mid return over [t-w, t]
  • r_H is Hyperliquid's mid return over the same span
  • gap = r_L - r_H, in basis points

Then look at Hyperliquid's forward return from t to t+h, for h in 250, 500 and 1000 ms.

On a two-minute BTC and SOL capture across the same four venues, at a 1000 ms window and a 500 ms horizon, the sign of the gap agreed with the next move 88 percent of the time and the mean forward move was -0.42 bps. That is not a paradox and it is not a coding error. It is a statement about the shape of the losses: the winners were small and frequent, the losers were large and rare, and the average is what pays the fees.
At a 250 ms window the correlation with the next 250 ms is actually negative, because that measure is reading the receiving venue's update granularity rather than a lead.

Then the cost screen, which is arithmetic rather than opinion. Hyperliquid's base tier is 1.5 bps maker and 4.5 bps taker. A maker entry with a maker exit and a taker fallback is 6.0 bps before slippage. Add a basis point of buffer and the hurdle is 7 bps. Of 88 anchors whose gap reached 5 bps, eight cleared 7 bps, and the mean forward move on those eight was +0.17 bps.

So: the structure is real, the edge is not, at least not in this capture on this network path. A high hit rate with a negative mean is the most common way a lead-lag backtest lies to its author, and both numbers come from the same trades. Quoting one is a choice.

Gold has three prices, and two of them are seven bps under the third

The same probe has a second mode that snapshots gold across rails. Same six seconds:

leg rail status bid ask premium ---------------- ------------------ ------ -------- -------- --------- OKX PAXG-USDT spot token (PAXG) ok 3,737.60 3,737.90 0.00 OKX XAUT-USDT spot token (XAUT) ok 3,737.10 3,737.20 -1.61 Binance PAXGUSDT USDT perpetual ok 3,735.00 3,735.01 -7.34 Hyperliquid PAXG perp, venue oracle ok 3,734.60 3,735.40 -7.36 

Premium is each leg's mid against the spot PAXG mid, in basis points. Both gold perpetuals quoted about 7 bps below the token that redeems for allocated London bars, for the entire window, on both venues, within 0.01 bps of each other. XAUT, a different issuer holding the same metal, sat 1.61 bps below PAXG.
That 7 bps is a basis, and a basis is not free money. It is what the funding rate exists to charge for, and whether it is reachable depends on the cost of holding both legs and the fee to cross each one. But it is a real, timestamped number that you can reproduce in eight seconds, and almost nobody publishing gold "arbitrage" content shows it next to the cost of the trade.

The London-New York leg is not in that table, because the COMEX contract and the LBMA lease rate are not on a free public socket, and a delayed print would poison a live cross-venue number. That bridge needs its own arithmetic, starting with the 73 hours between Friday's futures close and Sunday's reopen, when the most liquid gold price in the world does not exist while the tokens keep trading and the perpetuals keep marking.

What the probe does not do

  • It does not place orders, hold keys, or sign anything. There is no order router and no private endpoint in the repository.
  • It does not compute an edge. It prints the inputs: update resolution, spread, dispersion, basis.
  • It does not measure absolute one-way latency. Every timestamp is one machine's clock. The venue's own timestamp is kept beside it and never subtracted, because the difference contains a clock offset that is usually larger than the delay being looked for.
  • It does not claim six seconds is a study. It is a smoke test and a sample. Numbers move with the session, the region and the network path.

That last point is why the repository ships its capture file. replay prints the table above from a JSON file, so you can see exactly what the code does before you connect to anything, and you can diff your own capture against it.

Run it

python3 --version # 3.10 or newer, nothing to install python3 -m vprobe venue --symbol BTC --seconds 12 --save data/mine.json python3 -m vprobe gold --seconds 8 python3 -m vprobe replay data/mine.json # the shipped sample, no socket python3 -m unittest discover -s tests -t . # 54 tests, recorded frames, no network 

The code is MIT licensed and stays free: https://github.com/nvimuno-dot/venue-latency-probe

If you want the numbers in two pages

Two free tear sheets carry the headline results with the arithmetic behind them, no email required:

Two longer write-ups are paid, and I wrote both. A thirty-page report on gold's three pricing bridges and the six structures priced after costs is 49 USD. The measurement kit this probe is cut from, with 40 tests, a two-minute sample capture and a twenty-page guide to the definitions and their failure modes, is 100 USD. They are linked at the bottom of the repository README, and the probe is the free end of them.

If a number above looks wrong, open an issue with your capture file and the date. A lead-lag claim without a capture is an opinion, and that applies to this one too.