TWB The Weekend Basis

Methodology

How The Weekend Basis computes what it publishes. This page describes the calculation as it is actually implemented — if this page and the calculation ever disagree, the calculation is what produced the numbers and this page is wrong.

TWB holds no positions in covered assets

Summary. This is the short form the dashboard links to: what is measured, how it is judged, and what the numbers do and do not mean. The full methodology — every formula, threshold and limitation as implemented — replaces it here before launch.

What is measured

An xStock trades 24/7. The equity it tracks does not. Between the US close and the next US open there is no official price to mark against, so the token's price is free to wander — and the size of that wander is the number this project exists to measure.

Premium/discount is a token's deviation from its reference price in basis points. Weekend basis is that deviation summarised over one weekend window. Token above reference is a positive premium; below is a discount. Deviations are signed and always in basis points, at every magnitude.

A weekend window runs from the underlying's official closing price on its primary listing market (Friday 16:00 ET) to that market's next official opening price (Monday 09:30 ET) — the span in which no new official reference price is formed.

The window is bounded by the absence of a new official price, not by the absence of trading. The underlying already trades inside it on extended-hours and overnight venues, and additional overnight sessions — including the Nasdaq Night Session expected 2026-12-06 — do not shorten it, because they form no official opening or closing price. About a quarter of the window's hours already have the share trading somewhere in the US, and always have; what the window measures is the stretch with no new official price to mark against.

What the instrument is

Not equity. An xStock is a bearer debt instrument — a tracker certificate issued by Backed Assets (JE) Limited, a Jersey SPV, under an EU base prospectus, collateralized 1:1 by the underlying share held with regulated custodians under an independent security agent. It carries no ownership of the company and no voting rights.

Both halves are true together: the 1:1 collateral is real, and the wrapper is debt. The reference is still the equity's official close, because the certificate is written to track it. What changes is what a deviation from that close can mean: it can be liquidity, it can be the weekend effect this project measures, and it can be the market repricing Backed's structure. The headline number does not separate them and nothing here claims it does.

The headline number

The headline number is the weekend average — the time-weighted average deviation across the window. It is built from a mark every 60 seconds: one price observation per slot, taken from the midpoint of the token's order book. Time-weighted, not a simple mean, so a burst of quotes cannot outvote a quiet stretch of equal duration. Gaps in coverage — the slots with no usable mark — are never interpolated across. The dashboard and its CSV download label this number weekend twa.

The publication gate

Not every measurement is publishable. Each window is put through a fixed test — the gate — before any number from it reaches the dashboard, and the dashboard prints the verdict beside every row. A window is published only if its marks are fresh, its median spread is under a ceiling, and its headline clears a multiple of the token's own median spread. (The book is the exchange's live bid and ask for the token; the spread is the gap between them.) A deviation smaller than the spread it was measured across is not distinguishable from quote noise, so it is withheld — and the table shows the withheld row, with its spread, rather than dropping it. That is one of several ways a window is withheld: short coverage, a book over the ceiling, no usable quotes, or a failed unit check will each do it too. The row says which.

Reading the chart

The big chart's vertical axis is an index, not a price. It starts at 100 at the underlying's reference close and every level on it is the token's distance from that close, so 100.40 is 40bps above and 99.60 is 40bps below. It is deliberately not a dollar figure: a price at a stated instant, divided by the deviation published beside it, gives back the reference close exactly, and this page publishes no price of anyone's. The chart says so on itself, bottom right. The numbers on that axis are close to a share price for some names and nowhere near one for others — read them as an index or read the basis points beside them.

The marks are drawn to whole 5bps. Same reason, one level down: a run of marks against a fixed tick size is the exchange's own price grid divided by the reference close, and a grid divided by a number still points at that number. So the published marks are rounded coarser than the grid they came from. The step is five times the resolution this page prints at, so a mark moves by up to 2.5bps — whole printed basis points, not a fraction of one. The average, the peak and the spread band are not put on that step — they ship two decimals of a basis point, and this page prints whole ones from them, so those read the same either way. The chart does not. Its own highest and lowest marks land on the step — measured on the weekend of 2026-08-21, on the narrowest published name the high moves from 62.77 to 65bps — and the vertical axis autoscales to those marks, so its range moves on every published row: outward where the step rounds an extreme away from zero, inward where it rounds one in. The numbers printed up that axis move with it. On the same weekend one row's axis prints nine labels before the step and five after, at every chart width from 852 pixels up to the 1280 we measured — this page's text column is capped at 960 pixels, so that is the page as it ships — and other rows change labels at other widths. Through the flat stretches the line reads as a step function. That is a change of shape, and of the numbers on the axis, and not only of texture; it is what the step costs.

Reading the numbers

The table prints a deviation and the spread it was measured across; the panel that opens when you click a row prints five numbers. They are not five answers to one question.

The weekend average, against the spread beside it. The average is where the token spent the window; the spread is the width of the book it was measured in. Read them together or neither means anything: 40bps on a 2bps book is a move, 40bps on a 60bps book is the book. The gate applies that test, and the spread column and the chart's grey band are there so you can apply it yourself.

The peak, beside the average. The peak is the furthest the token got from the anchor — the official closing price the window starts from, which every deviation here is measured against — at any single mark, with the instant it happened; the panel prints it and the table does not. One mark and a whole window answer different questions — the peak says how far it went, the average says where it spent its time. A large peak over a small average describes one print rather than the window it sits in.

The Monday move, in three parts. A large snap-back means the equity did not validate the token's move — the equity reopened near Friday's close and the token was still away from it at the last mark before the open. Whether it converged after that is outside every number here: the last instant measured is the one before the equity's first print. A large equity gap means the token led: the stock reopened where the token already was, and the residual is small. On the weekend of 31 July, CRCLx closed 549bps below its anchor and CRCL itself reopened 580bps below — the token had it right, and the snap-back left it 33bps above CRCL's own open. Neither large means there was nothing to explain.

What "reopened" means here, and what it does not. Both the equity gap and the snap-back are measured against the first consolidated trade at or after 09:30 ET — not against the opening auction price. They are not the same number. Measured against NYSE's own auction feed, the two agree on 2 of 12 CRCL sessions, with a median gap of 20.5bps and a maximum of 100bps; the closing auction, run as the control, matches our stored close on all 12. Thirteen of the fourteen covered names are listed on Nasdaq or NYSE Arca, whose auction prints are not published free, so for those the size of this difference is unmeasured rather than small. It is a real limit on the two Monday legs and it is stated here rather than left to be discovered. The closing basis — the deviation at the last mark before the equity reopened — is unaffected: it never references the open.

What convergence here is, and is not. Through the weekend of 2026-08-07, two weekends had finalised — 14 published rows. The widest closing basis in them was 549bps below the anchor and the widest residual at the reopen was 33bps above the equity's open, both on CRCLx over the weekend of 2026-07-31. Those figures are pinned to that span; the dashboard and the CSV carry the current set. That is a record, not a mechanism. In an ETF the same record would be underwritten by creation and redemption: a wide enough premium pays someone to arbitrage it away. No such incentive reaches most holders here. Redemption at the issuer is KYC-gated and whitelisted, so a holder who bought the token on the secondary market has no such channel and nothing caps a premium except the next trade. In that same span the widest closing basis ended 33bps from the stock's own reopening print. That is a narrower claim than it sounds: the widest deviation in it was CRCLx's 644bps at a single mark 57 minutes before the equity reopened, and no reopening print was ever measured against that one — the Monday legs anchor on the last mark of the window, not on the peak. Both belong to those two weekends and neither is a standing record; the dashboard and the CSV carry the current set. These are observations of a market. The instrument promises none of it.

Independent recomputation

Every night a second, independent implementation — the auditor — goes back over the published numbers on a separate code path. It reprices the weekend from the exchange's trade prints and quotes rather than from the order book, over the same connection, and it takes the reference close and open from a second licensed vendor. It does not check every number the same way, and the difference matters.

A disagreement on the reference close, the window's boundaries, the arithmetic of each mark, the weekend average, the closing basis, the snap-back or the equity gap holds the whole row. A held row shows none of its numbers and none of its marks. Where the weekend average is on hold and an earlier window produced one that passed, the row shows that value instead, stamped with how long it has been held. Where there is no such value — the hold is on something else, or nothing has passed yet — the row shows no number at all rather than a substitute.

The peak is recomputed and compared, but a disagreement on it never holds a row: the furthest point of a trade feed and of an order book are not the same measurement. The spread is not recomputed, and neither is the coverage figure; how much of the window each side saw is reported beside the comparison. Some disagreements are not judged at all — a window the gate withheld, a comparison the auditor could make on too little of the window, a closing mark with no recent quote to measure against — and those hold nothing. The auditor does not strike over what it did not see.

The dashboard does not update unless every weekend in the file it loads has been reconciled. Before it writes anything, the nightly job that builds these pages asks the auditor's own record of its run two questions. First, did today's reconciliation finish, and with what verdict. Second — for each weekend in that file, the one in the dashboard's table and the previous one it summarises — has that particular weekend been reconciled since its numbers were finalised. Both have to answer yes.

If they do, the dashboard publishes, and anything the auditor quarantined is suppressed to its held value; a reconciliation that found a mismatch still publishes, because finding the mismatch is what quarantines the value behind it. If today's run did not finish, or compared too little of the day to stand behind it, or if either weekend has no reconciliation on record since it was finalised, it publishes nothing: the dashboard stays as it was and the timestamp above its table keeps saying when it is from. A stale page that says so is better than a fresh one nothing has checked.

The second question is asked because the first one cannot cover it. A day's reconciliation examines a trailing window of recent data, and the previous weekend stays in the file for longer than that — so there is a stretch in which the day's run is honestly clean while the older half of the file has not been looked at for some time. So the weekends the dashboard is showing are added to that trailing window rather than replacing it. A weekend is re-reconciled every night that either rule reaches it — which means every night it is on the dashboard, however far past the trailing window it has aged, and also on the nights before its numbers appear there at all, while they are still being finalised.

One clarification, because “reconciled” is doing specific work above. It means the auditor examined that weekend and stands behind what it examined — not that every number in it exists. A weekend where the price feed left gaps still reconciles: the gaps show up as the row labels and the missing numbers you can see in the table, which is the honest place for them. What the gate refuses is a weekend the auditor did not look at, or looked at and could compare too little of to say anything about.

Cadence

These pages are generated once a night, after the day's reconciliation, and the dashboard's header says when. Nothing refreshes during a weekend; the numbers you see are frozen at the time stamped in that header.

The other half of the header — market open, market closed · reopens …, weekend in progress — is about your clock, not ours. The dashboard's data file carries the upcoming session boundaries and your browser compares them against the time on your device, so the word is right whenever you read the page rather than right at 04:30 UTC, when the nightly run happens. It is not a claim that anything on the page is moving: nothing is, until the next nightly run. If the boundaries have run out — a nightly job that has not run for a fortnight — the header states no session at all rather than one it cannot support.