Sixty-one trades, sixty-one losses

This one is not about memecoins. It is about a trading engine I had the records for, and it is here because it fails in the same shape as everything else on this site: the part that was carefully designed was not the part that decided the outcome.

The trade

Funding arbitrage across venues. A perpetual futures contract charges a funding rate every so often, paid between longs and shorts, and the rate differs from one exchange to another. When it does, you can go long where funding is negative and short the same size where it is positive, hold both, and collect the difference. Your price exposure nets to roughly nothing. You are not betting on the coin. You are collecting a spread.

It is a real strategy and the arithmetic is genuinely attractive.

What the engine found

Over three days in May it identified 116 opportunities across ten instruments — BTC, ETH, SOL, XRP, DOGE, LINK, AVAX, BNB, ARB, HYPE — on four venues. The median opportunity was worth 9.2% a year. The best was 87%.

Those are not fantasy numbers. They came from thirty-one thousand funding snapshots read off the venues themselves.

What actually happened

It opened 61 positions, $710,000 of notional, every one of them paper.

Sixty-one lost money. Not fifty-nine, not "most". Every single one.

Total realised: −$1,678.50. Total funding collected across all sixty-one positions: $7.77.

Why

The median position was held for 1.3 hours.

Opening one costs about 24.5 basis points across the two legs, in fees and modelled slippage, and closing it costs the same again. Call it 49 basis points for a round trip.

A 9.2% annual spread earns about 0.025% a day. To cover 49 basis points you have to hold the position for nineteen days.

The engine held for eighty minutes. That is not a strategy that underperformed. That is a nineteen-day trade being run as a one-hour trade, and it was mathematically certain to lose before a single order was placed. The realised losses confirm it in the least ambiguous way possible: every figure in the records is exactly the modelled round-trip cost of the position. $24.50 on a $10,000 position, $50 on a $20,000 one. Nothing else happened. The market never got a vote.

There is a second mechanism underneath the first. Funding does not accrue continuously — it is paid at fixed intervals, and the venues disagree about when. Three of the four pay every eight hours. The fourth pays hourly. A position that lives eighty minutes usually spans no funding payment at all on the eight-hour side and perhaps one on the hourly side, which is why sixty-one positions with a median 9.2% spread collected seven dollars and seventy-seven cents between them. The legs were hedged on price and not hedged on funding, because they were not held long enough for funding to exist.

What this is not

Paper trading, throughout. No real money was committed and none was lost. The losses are what a cost model says the trades would have cost, not what a market charged, and the eight-basis-point slippage assumption is an assumption — real fills could be better or worse.

Three days is a very short window. Five positions were never closed, and three rows carry a close timestamp earlier than their open, which is a clock or recording fault rather than a trade.

And an engine that never went live is not a strategy that was disproved. It is a strategy that was costed.

The part worth keeping

The costing was the whole return on the exercise. Nothing was risked, the arithmetic was allowed to run, and it produced a number — nineteen days — that made the decision by itself. Whether the spread was real was never the question, and it was real. The question was whether the holding period the engine was built around could survive its own transaction costs, and that could have been answered on paper in an afternoon.

This desk keeps arriving at the same place from different directions. Four launches said the coin was not the variable. Seventy-three tracked projects said attention arrives after the money, not before it. Sixty-one paper trades say a sound edge held for the wrong length of time is not a small loss, it is a guaranteed one.

The common thread is that in each case the thing being optimised — the name, the marketing, the signal detection — was downstream of a constraint nobody had priced. The cheapest work available is finding that constraint before spending anything, and it is almost always arithmetic rather than insight.