We told people weekends cost more. Ten days of measuring says we were wrong.

On 31 July we published a rule. A tokenised stock is quoted by a market maker who hedges the real share on the real exchange; when the exchange is shut they cannot hedge, so they widen. We had the receipt: Broadcom cost 0.11% to enter on a Friday and 1.99% the next morning.

We then started sampling four times a day and writing every reading down. There are twenty-seven of them now, from 1 to 10 August 2026. The rule does not survive them.

The weekend is not the expensive window. For most of what we sell it is the cheapest one. And Broadcom, the asset the rule was built on, turns out not to care what day it is at all.

What is behind these numbers#

Four samples a day, at the open, at midday, at the close and overnight. Each one quotes every asset in the registry against live markets at $100, $1,000 and $5,000, and records the price impact at each size. The file is src/data/cost-history.json in the open-source repo and the sampler is scripts/measure-cost.mjs.

Twenty-seven samples split three ways:

• New York open — eight samples
• Overnight — the weekday hours when the New York equity market is shut, eleven samples
• Weekend — eight samples

Everything below is a median of the price impact on a $1,000 order. Medians rather than averages, because a single bad reading should not be able to write the headline — which is precisely the mistake this article is about.

The worst time to buy is a weekday night#

Not the weekend. The overnight window, when New York is shut but the world is otherwise working.

• Apple: 0.49% open, 0.68% overnight, 0.33% weekend
• Microsoft: 0.38% open, 0.53% overnight, 0.46% weekend
• Alphabet: 0.21% open, 0.42% overnight, 0.11% weekend
• Amazon: 0.42% open, 0.48% overnight, 0.07% weekend
• Ondo's Treasury token: 0.16% open, 0.32% overnight, 0.15% weekend

Five for five, overnight is the worst of the three. It is the window nobody warns you about, because the folk theory has one boundary in it — market open, market shut — and the data has two.

The weekend is the cheap window, not the dear one#

For the large single names the weekend is where the tightest quotes are.

• Amazon: 0.42% while New York trades, 0.07% at the weekend. Six times cheaper.
• Alphabet: 0.21%, then 0.11%.
• Apple: 0.49%, then 0.33%.
• Nvidia: 0.14%, then 0.10%.

It holds at size. At $5,000, Amazon runs 0.59% during the session and 0.23% at the weekend; Alphabet 0.31% and 0.18%. This is not a small-order artefact.

We are not going to pretend we know why. A plausible story is that weekend flow is thin in both directions and the makers who stay quoted are the ones who have chosen to be there, while the overnight window catches the shift change — real positions being adjusted against a market that has closed and not yet reopened. That is a story, not a measurement, and this site's whole point is the difference.

The index product does the opposite#

• QQQx: 0.01% while New York is open, 0.08% overnight, 0.18% at the weekend.

Seventeen times worse on a Saturday than during the session. QQQx is the one asset that behaves the way the rule said everything would.

SPYx barely moves at all — 0.03% in every window — which is its own kind of answer about where the depth on Solana actually is.

Broadcom does not care what day it is#

This is the part that took the rule apart.

Broadcom's entry cost, across twenty-seven readings, is one of two numbers. It is either somewhere between 0.22% and 0.49%, or somewhere between 1.76% and 2.21%. Fourteen readings in the expensive state, thirteen in the cheap one.

Nothing has ever been measured between 0.5% and 1.5%. Not once.

It flipped state eight times in ten days, and it flips inside every kind of session. On 2 August it read 2.19% at 06:04 and 0.23% seven hours later, both on a Sunday. On 5 August it was 1.76% overnight, 2.12% through the session, and 0.23% by 22:00. On 7 August it went 0.23% at 00:59 and 2.21% four hours later, with the market shut for both.

That is not a spread widening when the hedge goes away. That is one venue appearing and disappearing, and a router that either finds it or does not.

Which makes the rule we published from a Friday and a Saturday exactly what our own notes warn against. From AGENTS.md, written before any of this was sampled:

Cost is a distribution, not a number. One probe cannot tell a bad asset from a bad hour.

We wrote that down and then published a rule from two probes anyway.

What a buyer should do with this#

Nothing, and that is the useful part.

Do not time the clock. The relationship is real for the index and inverted for the single names, and for Broadcom there is no relationship at all — only a coin that lands two ways. Any rule simple enough to remember is going to be wrong for a third of what is on this site.

What works instead is what the order screen already does: it quotes every leg live, at the size you chose, at the moment you are looking. If Broadcom is in its expensive state you will see 2.2% on the screen before you sign anything. If it has flipped, you will see 0.3%. That number is measured, not modelled, and it is the only one worth acting on.

The one place the clock matters is patience. If you are looking at Broadcom at 2.2%, the readings say it will probably be back near 0.25% within a day or so. Nothing forces you to buy the expensive draw.

What we changed#

The FAQ answer on our previous article said the same token costs more on a weekend, and cited the Broadcom Friday-to-Saturday jump as the reason. That answer has been rewritten. The jump was real; the explanation was not.

We would rather correct it in public than leave a tidy rule standing over messier data. The measurement is the product. It is not worth much if it only ever confirms what we already said.

Limits worth stating#

Twenty-seven samples over ten days is enough to kill a rule and not enough to build one. Eight readings during market hours is thin. The overnight bucket mixes European morning with American night. One probe per sample means each number is a single draw from whatever Jupiter's routing found at that second. US public holidays are filed as open, which is a known small error in the sampler.

We keep sampling. The file only gets longer, and every reading in it is timestamped and public.

The full universe measurement is free as JSON at /api/universe, CORS open, free to use with attribution. The cost history is in the repository. If you write about tokenised equities and want the numbers, take them.