The idea
Some exchanges pay you to hold a position that others charge you for. Hold both sides at once and you collect the difference while taking no real view on the price. It is one of the few genuinely non-competitive trades left for a small operator: nobody has to be fastest, the payment just accrues.
We built a measuring instrument for it — no orders, ever, only observation — and set it running for thirty days across several venues. We also wrote down, in advance, the five conditions the results would have to meet. That is the house rule: the criteria are frozen before the data arrives, so that nobody gets to be clever afterwards.
The result
Ten thousand simulated round trips. The verdict:
- Collected in rate differences: about five dollars per trade
- Paid in trading fees: twelve dollars per trade
- Net: about seven dollars lost, ten thousand times over
Fees came to more than twice everything the strategy earned. Two of the five conditions failed outright, and the rule says any failure is a kill. So it was killed.
Worth noting what did not go wrong: no individual position was dangerous. The worst single trade lost twelve dollars. There was no blow-up, no bad market, no surprise. The strategy simply bled to death through the turnstile.
Then we did the arithmetic, and it got worse
The rate differences are paid out slowly, drip by drip. The fee is charged all at once, when you go in and when you come out. So the only question that matters is: how long must you hold before the drip covers the toll?
We had never asked it. So we asked it afterwards:
| If the rate difference is… | You must hold it for… |
|---|---|
| the minimum our rules allowed | five and a half days |
| three times that | nearly two days |
| six times that | about a day |
And our rules permitted exiting after eight hours.
Read those two facts together. Our entry condition said "only trade when the difference is at least this big." Our exit condition allowed leaving after eight hours. At that combination, a trade collects roughly seven cents against twelve dollars of cost.
There was no market in which these rules could have made money. Not a good month, not a bad one. We had written a set of instructions that guaranteed a loss and then spent thirty days carefully measuring how much.
The part that stings
The experiment was well run. Real data, honest fees, no orders placed, no capital risked, criteria frozen in advance, a bug caught and fixed midway. Every piece of process worked.
And all of it was in service of a question that a two-line calculation had already answered. We validated the strategy against the market and never validated the rules against arithmetic.
This is the third time in a month we've found the same shape of mistake. One gate would have declared a dead strategy alive because it measured a gap without asking whether the gap was already explained. Another would have triggered on a missing data field. Now this one — thresholds that could not clear their own costs.
The lesson we keep re-learning, more expensively each time: a decision rule is code, and code can be wrong. We test our strategies ferociously and accept the rules that judge them without a second glance, as though writing something down in advance makes it correct. Committing to a threshold early protects against fooling yourself later. It does nothing about being wrong at the moment you write it.
What we changed
One line, added to the process: when you write an entry threshold, compute what it must earn to clear the cost of trading — before you collect any data. If the threshold cannot pay for the transaction, there is no experiment to run.
What survives
Exactly one thing. All those fees came from demanding immediate execution. Place patient orders instead — wait for someone to come to you rather than crossing the spread — and the cost drops by roughly five-sixths. At that level, a trade needs a few hours instead of several days, and a few hours is inside what we actually observed.
That version was never built. It is now the only condition under which this idea reopens, and it is a rebuild rather than an adjustment.
So the month ends with a strategy in the graveyard, a rule added to the process, and one clearly-specified door left open. Not the outcome we wanted. But the failure was free, it was found in thirty days rather than thirty thousand dollars, and the thing we learned applies to every strategy that comes after — which is more than most winning trades can claim.