On Friday at 17:00 New York time the most liquid gold price in the world stops printing. It does not print again until Sunday at 18:00. Seventy-three hours in which the contract that sets the marginal price of a 100-ounce bar is closed, the vaults that lend it are quiet, and the only people quoting gold are the ones running on-chain books.
Gold does not stop existing during that window. Three kinds of venue keep answering the question "what is gold worth" anyway:
- a token on a crypto exchange that claims one allocated London bar behind each unit,
- a perpetual future on a large centralized derivatives venue, marked against its own index,
- a perpetual future on a venue that marks against a token price.
Each of the three returned a number every six seconds through the whole weekend. The numbers were not the same number, and the ordering between them was not random. That ordering is the subject of this post, because it is about seven basis points, and seven basis points is exactly the size of the thing most people cross without noticing.
Six seconds of three gold prices
The probe is about 300 lines of standard-library Python with no key and no account. It subscribes to public top-of-book on four venues on one clock and prints spread, mid and basis. Around eight seconds of its gold mode:
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, from the same capture. The spot token is the reference because it is the one leg you can redeem for metal: 3,737.75 at the mid, against 3,735.00 and 3,734.65 for the two perpetuals.
Read the last column twice.
Two perpetuals, on two venues with different matching engines, different customer bases, different funding schedules and different index construction, quoted 7.34 and 7.36 basis points under the redeemable token. They agreed with each other to within one hundredth of a basis point, and disagreed with spot by seven.
That is not a coincidence and it is not a mispricing you can pick up. A structure that two independent venues reproduce to 0.01 bps is a structure with a price, and a price has a payer. The useful question is not "can I buy the cheap one and sell the expensive one" but "what exactly is that 7 bps charging me for".
Three answers to the same question, and only one of them is gold
During the blackout the futures market is not there to arbitrage against. So each venue has to invent a mark, and the invention differs by design. There are three common designs:
A frozen index. The venue takes the last index value from before the close and holds it. The perpetual keeps trading against a mark that cannot move, so the book develops whatever basis the crowd wants, and funding is the only thing pulling it back toward a level that stopped being real on Friday afternoon.
A bounded order-book average. The venue computes the mark from its own book, fenced inside a band around the last known price. The mark can move, but only as far as the fence allows, and a weekend that gaps two percent will sit pinned at the edge of the band until the index wakes up.
A token price. The venue marks against an asset that trades 24/7 on someone else's book. The mark is continuous and it is liquid, and it also inherits every weekend-specific behaviour of a market where the usual participants are asleep.
None of these is the price of gold. All three are prices of gold-shaped instruments, and the 7 bps in the table is the market's running estimate of the difference. That is why the two perpetuals could agree to a hundredth of a basis point: they are both marking the same continuous thing, seven bps below the token that redeems.
And the token itself is not free of tolls. It has a redemption schedule and a creation schedule, and premium to spot is bounded by the cheaper of the two plus the cost of the metal. A basis inside that band is a normal operating state, not a signal.
The part that costs money
Here is where the weekend story turns from interesting to expensive for anyone who acts on it.
Suppose you see a 7 bps discount and decide to own it: long the perpetual, short the token, wait for convergence on Monday. Your gross is 7 bps. Now itemise what you paid to collect it.
Crossing the perpetual book costs the taker fee plus the half-spread. On a venue whose 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, which is why the cost screen I use everywhere in this work is 7 bps: six of fee, one of buffer.
So the trade that looked like seven bps of free structure is roughly six bps of fee and one bps of compensation for being right about the direction and the timing of a convergence that will happen on the other side of a weekend, at a moment chosen by an exchange calendar.
If you take both legs (perp and token), the spread you pay doubles. If you fund the short over the weekend, funding is charged at whatever interval the venue chose, and at least one major venue widened that interval during 2026, which silently doubles the error in any carry model calibrated before the change.
The reason to write this down is not to warn you off gold. It is that the seven bps is the visible part of the trade and the fee is the invisible part, and the two are the same size. Most descriptions of a gold basis stop at the first number.
The same arithmetic, one rail over
The identical failure shows up in crypto perps, where the numbers are cleaner and only slightly more embarrassing.
On a two-minute capture of BTC and SOL across Hyperliquid, Binance USD-M, OKX and Bybit, at a 1000 ms window and a 500 ms horizon, the sign of the cross-venue gap agreed with Hyperliquid's next move 88 percent of the time. The mean forward move over those same anchors was -0.42 bps.
Eighty-eight percent right, and poorer on average. That is not a bug in the calculation; it is the shape of the distribution. The wins were small and frequent, the losses were large and rare, and the average is the number that meets the fee schedule. Of 88 anchors whose gap reached 5 bps, eight cleared the 7 bps hurdle, and the mean forward move on those eight was +0.17 bps.
The same probe shows why the low end of that measurement is not even measuring what it claims. In six seconds of BTC, Hyperliquid printed 11.0 top-of-book updates per second with a median gap of 75 ms; Binance USD-M printed 376.5 per second at under a millisecond. Four mids ended 1.17 bps apart. If the receiving book only prints a new touch every 75 ms, a 250 ms "lag" is the venue's tick granularity wearing a stopwatch.
Two markets, one lesson: the number that looks like an opportunity is routinely the same size as the number that pays for it, and the only way to know which one you are looking at is to put both on the same line.
Run it yourself
The probe ships with the capture it printed, so replay gives you the tables above before you open a socket. Nothing to install, no key, no order router, no private endpoints.
python3 --version # 3.10 or newer
python3 -m vprobe gold --seconds 8
python3 -m vprobe venue --symbol BTC --seconds 12 --save data/mine.json
python3 -m vprobe replay data/sample_snapshot.json
python3 -m unittest discover -s tests -t . # 54 tests, recorded frames, no network
MIT licensed and free: https://github.com/nvimuno-dot/venue-latency-probe
If a number above looks wrong, the useful reply includes your own capture and the date, and the repository makes that easy to produce.
Where the rest of the arithmetic lives
Two free tear sheets carry the formulas and the tables, no email required:
- The 73-Hour Gold Gap and EFP Tear Sheet, on the three weekend marks, the reopening move against a round trip, and what an EFP costs per ounce: https://baobab-field-press.shopfa.st/products/gold-gap-efp-tear-sheet
- Why 88% Accuracy Still Loses 0.42 bps, on the hit rate, the mean, and the 7 bps screen: https://baobab-field-press.shopfa.st/products/leadlag-paradox-tear-sheet
The longer work is paid, and I wrote it. The gold report is 30 pages on London, New York and on-chain pricing, with six structures priced after costs, at 49 USD: https://baobab-field-press.shopfa.st/products/three-bridges-of-gold. The measurement kit this probe is cut from, with 40 tests, a two-minute capture and a 20-page guide to the definitions and their failure modes, is 100 USD: https://baobab-field-press.shopfa.st/products/hyperliquid-leadlag-research-kit. Both together are 129 USD: https://baobab-field-press.shopfa.st/products/market-structure-bundle.
The probe is the free end of all of it and stays free.
The question I would actually like answered
If you hold a gold position over a weekend on a venue that freezes its index at the Friday close, you are not long gold. You are long one venue's answer to the question, and that answer was written before the news arrived.
So: what does the mark do on a Saturday when the token moves two percent? I have captures from quiet weekends and I would rather have one from a weekend that moved. If you have a recording of a weekend gap on a perp whose index was frozen, I would like to see it, and I will test it against the same arithmetic.
Free tear sheets, no email and no account: the gold side is https://baobab-field-press.shopfa.st/products/gold-gap-efp-tear-sheet and the lead-lag side is https://baobab-field-press.shopfa.st/products/leadlag-paradox-tear-sheet. The probe, the capture and every table in this post are at https://github.com/nvimuno-dot/venue-latency-probe and stay free.
Top comments (0)