The fee truth: two of our three trading paths were negative after costs

The finding that ruined our week

Two of our lab’s three trading paths were net negative after fees. Gross, some of them were in the black.

That is the whole story, and it is more uncomfortable than a loss from bad forecasts. A wrong signal can be repaired. A cost structure that structurally consumes the return turns even correct signals into a minus — it works silently, evenly, and regardless of whether the market moves our way.

So we did not examine the strategy. We examined the execution. A transaction-cost analysis does not ask whether a trade was right. It asks what the trade lost on the way from signal to fill.

How this number is built

So the figure stays traceable, here is the arithmetic in the open: per path we take the settled fees from the fill data and divide them by the gross profit of that same path over the same period. Both quantities come from settled records — not from a simulation, not from an estimate. It is a division of two line items the venue itself signed off on.

What this number is not: externally audited. It is our measurement of our system, and nobody outside can recompute it. That is why we disclose the method rather than only the result, and why a separate section below deals with exactly where that same measurement turns unclean. Anyone who does not believe our numbers should at least be able to check how they came about.

Path A gave up roughly 2.25 points of fees for every point of gross return. Path B roughly 1.18. Both paths could have been right on every single trade — the result would have stayed negative.

One thing matters for reading this number: the ratio contains settled fees only. Slippage and funding are not in it. The bars are the lower bound of the friction, not its full extent.

The anatomy of friction

Friction is not a line item. It is a chain, and every link looks harmless on its own.

The taker fee. Our executions were 100 percent taker: we take liquidity out of the book instead of offering it, and pay the full rate of roughly 6 basis points per fill. That sounds like nothing. Six basis points are six hundredths of a percent.

The slippage. Between the price a signal means and the price a fill gets, there is a gap. We estimate it at roughly 6 basis points. It does not come from malice but from physics: the book moves while the order is in transit.

The funding. Anyone holding a position across the settlement intervals pays (or receives) funding. At short holding times that is a rounding error. At longer ones it is not.

Here is the part that mental arithmetic skips: a round trip is two fills. In and out. The fee applies twice, the slippage applies twice. Roughly 6 basis points of taker plus roughly 6 basis points of slippage, times two sides — the position starts twelve basis points down and has to make up that deficit before the market has formed an opinion at all.

A single round trip absorbs that. The single round trip is never the question. The question is how often.

The win-rate illusion

The win rate is the most popular number in trading and the least informative. You can win the majority of your trades and still lose.

Two conditions are enough. First, payoff geometry: if the losers are bigger than the winners, a majority of wins does not outweigh a minority of losses. We measured exactly that elsewhere in the lab and published the numbers in our lab story about the dismantled strategy. Second, frequency: every round trip nibbles off its friction, regardless of outcome.

The result is a product, not a sum:

Geometry × frequency × fee = the real result.

Optimising one of the three factors optimises nothing. An excellent win rate with poor geometry and high frequency is a well-camouflaged loss machine — and it feels better to trade than a mediocre win rate with clean geometry, because the little hit of success arrives more often.

This is not an observation we invented. Barber and Odean evaluated 66,465 households at a large US discount broker over six years: those who traded most earned 11.4 percent a year while the market returned 17.9 percent. The title of their paper is the summary — trading is hazardous to your wealth. The mechanism there is ours: not the selection, the turnover.

That part of the cost only appears at execution is equally well described. Almgren and Chriss show that market impact splits into a permanent and a temporary component, and that a trade-off exists between execution speed and cost: take fast, pay impact; execute slowly, carry market risk. Neither model contains a free execution.

What our own measurement cannot do

An admission belongs here, and it belongs here before somebody else finds it.

Our slippage measurement is imprecise in one place. We derive the intent price from notional divided by size. For high-priced coins in tiny position sizes, that division rounds coarsely — the derived price is then not a measurement but an artefact of the rounding. Concretely: individual per-symbol values between minus 150 and minus 270 basis points sit in our data. They are not real. They are the noise of the method, not the behaviour of the market.

We could leave them out. We leave them in and write down what they are.

The clean route would be a reference price at the moment of the signal — the mid price you measure the fill against. We lack a connector for that. This is an open point, not a solved one, and therefore we deliberately draw no recommendation from the slippage values. A wrong number with a recommendation on top would be worse than no number at all.

And here is why the core claim stands anyway: the slippage is not inside the fee share. The ratio of 225 and 118 percent is computed from settled fees and gross profit — both signed-off line items, not estimates. The unclean quantity sits outside the number that carries the headline. Folding the estimated slippage in would make the result look worse, not better. We leave it out because it is an estimate, and we accept the less favourable case for our own thesis.

So what holds: the fee share of gross profit and the net edge after costs. What does not hold: any claim about the size of our slippage per symbol.

And the third path? It was marginally positive after fees. We do not sell that as an edge, because the sample covers roughly 100 fills, about 15 round trips. Fifteen round trips prove nothing. They hint that the friction was bearable there, and nothing more.

The research we cite has limitations too, and we will not hide them. First, so the roles are clear: the two studies do not prove our numbers and are not meant to. Our numbers are the exclusive element; the studies are the evidence that the mechanism behind them is neither new nor our special case — it has been described for decades, in other markets, with other instruments, with the same outcome. Barber and Odean studied clients of a single US discount broker between 1991 and 1996, directly held equities only — no crypto markets, no 24/7 trading, an entirely different fee world. Their study observes, it does not experiment; overconfidence is an offered explanation there, not a proven causal link. Almgren and Chriss, in turn, supply a theoretical model with a deliberately simple linear cost approach, built for large institutional orders. The magnitude of their market impact does not transfer to our lab’s tiny fills. Both works support the mechanics of our findings, not their numbers.

The lesson

Four things we take away, and they hold regardless of what somebody trades.

Calculate net, not gross. Every strategy has to be worked through after all costs — fees, slippage, funding. A gross curve is an advertisement for a strategy, not a statement about it.

Frequency is a cost factor, not proof of diligence. Fee drag scales with the number of rounds. Fewer, larger trades beat many small ones once friction is involved. That is the most uncomfortable sentence for anyone who enjoys the act of trading — and it is the same finding the high-turnover households produced in Barber and Odean’s data.

Maker or taker is not a detail. Anyone who only ever takes liquidity structurally pays the full rate. Anyone who offers it waits — and carries the risk of never being filled. That trade-off should be a decision, not a side effect of the code. With us it was a side effect.

Ask one single question of every profit screenshot: gross or net? If the answer does not come, or dodges, it has been answered. People who know their costs name them gladly, because it makes their numbers more credible. Telling clean numbers from pretty numbers is, incidentally, the same skill you need for spotting crypto scams.

We write this down because our lab stands in the workshop, not in the display window. More calculations of this kind, including the unpleasant ones, live in the lab section. What else we disclose about these numbers — method, limits, open points — sits in our transparency overview.

The fees taught us nothing about the market. They taught us something about ourselves.