Cost to Hold · No. 0003 · Thursday 13 August 2026 · five minutes
Everything is down about a point and a half and the funding underneath it went in four directions at once. Today: why a confident number is the dangerous kind, a filing that would put ETF holders on the earning side of the same trade you are paying for, and three venue-assets that have snapped back to the default rate and stopped saying anything at all.
You cannot know tomorrow's price. You can know today's funding.
Nothing should be near the order button that cannot say why. Not because the tools are bad — because of two specific things they do to numbers, and one thing they do to your record.
A model that searches is a trial-count machine. Point one at price history and it will test ten thousand configurations and hand you the best. That is not a better edge than testing ten and keeping the best. It is a better-looking artifact, found faster, and the number attached to it means less in exact proportion to how hard the search worked. The count of what was tried is the only thing that makes the result interpretable, and an automated search is precisely the setting where nobody writes it down.
Underneath that sits a subtler failure, and it is the one that should decide where you let a language model near any of this.
A language model does not calculate. It predicts what a number ought to look like, given everything it has seen. That is a different operation from arithmetic and it fails in a specific direction: the output is plausible by construction — right digits, right range, right units, and sometimes simply wrong. A number that looks wrong gets caught by anybody. A number that looks right is the one that gets published.
Perpetuals are close to the worst instrument you could pick for that failure.
Ask what it costs to hold ether perp for a month and any confident answer is already suspect, because the cost is a rate multiplied by how often it is charged and only the rate is published anywhere. Bybit's own documentation says settlement frequency adjusts automatically when funding reaches its limit, switching to hourly and then reverting to two, four or eight hours "depending on market conditions, without prior notice." Bybit The multiplier in that sum is not a constant, it is not announced, and it is not in the feed. Anyone quoting an annualised figure across venues has guessed at it, whether they know it or not.
The inputs also move faster than any static answer survives. On Tuesday ether funding ran from 3.44% annualised on OKX to 8.13% on Binance. By Wednesday morning OKX was 4.20% and Hyperliquid had climbed to 10.95%. Solana went further: Binance paid longs −2.11% and charged them +3.83% a day later, while Bybit did the exact reverse. A number that was right on Tuesday was wrong on Wednesday, and nothing about it looked any different.
Which is why the tape below marks Binance and Bybit funding assumed 8h, and says so every day. We cannot confirm the interval from the feed, so the arithmetic refuses to pretend. A labelled assumption is worth more than a confident number, and it is the one thing a model will never volunteer.
So the question is not whether to use the tools. It is which half of the work they get. Hand one the arithmetic and you have added a confident guess to your process. Make it run deterministic code and report what the code returned, and you have added something that can be checked line by line and re-run tomorrow against the same inputs to get the same answer. The model proposes; the arithmetic disposes. Same shape as the trial count — a search may generate whatever it likes, so long as nobody trusts the generator and something mechanical does the killing.
Then the part that touches the record. An order you approved carries an intent. An order a model placed carries a timestamp and a fill. Six months later, working out whether the edge decayed or the venue simply got more expensive, one of those is evidence and the other is a receipt.
Automation is fine nearly everywhere. It stops being fine in the two places that destroy the audit trail: the search that produces the strategy, and the decision that commits capital.
The fair objection
If a venue can change its settlement frequency without notice, and funding can invert across two exchanges overnight, how is a backtest reliable at all? It was run against a cost structure that no longer exists.
That is correct, and it is not a flaw to be engineered away. A backtest cannot anticipate a regime change and should never be asked to.
What it can do is tell you how far the cost has to move before the edge stops existing. Not did this work but what does this survive — the break-even cost level, the share of gross edge that fees and funding consume, whether it still holds when every fill crosses the spread. Those are not predictions. They are tolerances, and they stay true when the environment does not.
Put that against the tape and the question answers itself. If an edge dies above roughly 6% annualised funding, and four venues are printing anywhere from 3.42% to 10.95% on the same asset, that is not a strategy problem. It is a venue selection problem, and it was solvable this morning before anything was placed.
Which is the part most people have backwards. You cannot know tomorrow's price; that is the whole reason this is difficult. But you can know today's funding across four venues, and today's fee tier, and what your interval actually is. The variable everyone treats as fixed is the one you can observe in advance. The variable everyone tries to predict is the one you cannot. Almost all the effort goes into the second.
Also today
Fidelity is adding staking to its spot ether fund. The filing for the Fidelity Ethereum Fund says it could stake up to 100% of holdings, keeping 85% of the staking rewards for the fund and taking 15% as fees, beginning "as soon as practicable" after the prospectus clears. Cointelegraph
That widens the gap this newsletter keeps pointing at, in the most direct way available. The ETF holder would earn a yield on ether. The perp holder with the same directional exposure pays for it — today anywhere from 2.10% annualised on Binance to 10.95% on Hyperliquid, an eight-and-a-half-point spread on the same asset before a staking yield is added to the other side. Perps buy you leverage, immediacy and no custodian, and those are worth paying for. It is worth knowing what you are paying.
Three venue-assets have snapped back to the default and stopped saying anything. Hyperliquid's ether and solana funding and Bybit's bitcoin funding are all sitting at exactly 10.95% annualised this morning — which is not four decimal places of coincidence, it is the 0.01%-per-8h baseline. Solana on Hyperliquid got there by jumping 9.6 points in a day. A rate at baseline is not the market agreeing on anything; it is the premium having gone to zero, leaving the default rate doing all the work. It looks exactly like a stable number until you check what it is pinned to.
And the macro shrug is the story. July's print landed exactly on consensus, which trimmed rate-hike odds from roughly 48% to 40% and moved digital assets almost not at all: bitcoin closed Wednesday at $63,768, up 0.28%, ether at $1,886.51. The read is that traders had already priced the Fed's path and have moved on to watching regulatory and corporate moves instead. Rio Times
Which is exactly what our own tape showed. Price did nothing on the print. The carry moved six points on one venue and not at all on another, and none of that was visible on a chart.
The tape
Price and 24h change
| BTC | ETH | SOL | |
|---|---|---|---|
| Mark | $63,359 |
$1,876.20 |
$75.57 |
| 24h | −1.25% |
−1.92% |
−1.47% |
Funding, per 8h and annualised
| BTC | ETH | SOL | |
|---|---|---|---|
| Hyperliquid | +0.0082% · 8.95% |
+0.0100% · 10.95% |
+0.0100% · 10.95% |
| Binance | +0.0053% · 5.79% |
+0.0019% · 2.10% |
−0.0020% · −2.18% |
| OKX | +0.0078% · 8.52% |
+0.0061% · 6.65% |
+0.0014% · 1.56% |
| Bybit | +0.0100% · 10.95% |
+0.0028% · 3.03% |
+0.0028% · 3.06% |
Hyperliquid settles hourly; its rate is shown multiplied by eight so the column compares. Binance and Bybit are read from CoinGecko rather than directly, and their interval is assumed to be 8h — see the venue note below.
Open interest, basis, and 24h change
| BTC | ETH | SOL | |
|---|---|---|---|
| OI, Hyperliquid | $2.57B +4.2% |
$1.62B −3.5% |
$0.38B −3.5% |
| OI, Binance | $6.99B +1.5% |
$4.43B +0.5% |
$0.65B −1.6% |
| OI, OKX | $2.06B +0.9% |
$1.29B −1.8% |
$0.23B +1.3% |
| OI, Bybit | $3.74B +1.4% |
$1.39B −1.6% |
$0.48B −4.4% |
| Basis, Hyperliquid | −4.4 bps |
−3.2 bps |
−4.2 bps |
| Basis, OKX | −3.9 bps |
−4.1 bps |
−5.3 bps |
Bitcoin open interest rose on all four venues while price fell — positions being added into weakness rather than closed out of it. Ether and solana went the other way almost everywhere.
Captured 2026-08-13 11:36 UTC · raw snapshot
Venue watch. Binance returns HTTP 451 and Bybit HTTP 403 to direct
requests from this location — geographic blocks, not outages. Their columns come
from CoinGecko and are marked as such, so the coverage gate records this issue
as degraded: two venues first-party, two second-hand.
One question
If something automated placed your last ten orders, could you say why each one was placed — from your own records, not from memory? Reply and say. I will publish what comes back, without names.
The cost corner
Funding is a cost of holding, not a fee on trading — which is why it appears nowhere in the number most people call their P&L.
The Founder No. 01 — Your exchange is deleting your trading history
Not investment advice. Nothing here predicts anything. Numbers are pulled from venue APIs at the time stamped above and may be revised. Third-party figures are linked to their sources.
Written by Joe DeFeo from inside a live systematic book.
Provenance
Captured 2026-08-13T11:36:38Z
Venues 4 of 4 reachable
Gates DEGRADED · 3 flags
Snapshot /daily/2026-08-13/data.json — append-only, never edited
- coverage: 2 of 4 venues read first-party. Binance and Bybit are filled from CoinGecko and labelled in the tape.
- third-party funding interval assumed 8h — the feed does not state it.
- liquidations unavailable — no free venue endpoint serves a 24h total.
Every number above can be checked against that file. If one is later revised the correction runs here and at the top of the next issue — the snapshot itself is never rewritten.