Most backtests model costs as one number. A fee percentage, applied at entry and exit, maybe a fixed slippage assumption if the author was being careful.
What you actually pay is a stack of separate charges. They arrive at different moments, they scale with different things, and they vary a lot depending on what you trade. Collapsing all of that into a single percentage is one of the main reasons a strategy performs worse live than it did in testing.
Here is the stack, in the order you pay it.
One round trip
Take a concrete trade. You buy one unit of something at around $64,000 and sell it later. Assume a perpetual contract, since that is where most crypto volume sits.
The exchange fee. Say five basis points as a taker. That is $32 going in and $32 coming out. Everyone models this one.
The half-spread. The book shows a bid at $63,995 and an ask at $64,005, so fair value is $64,000. Buying immediately means paying the ask, and you are $5 behind before the position has done anything. Selling later at the bid costs another $5. Nobody sends you an invoice for it, which is why it disappears from cost models so often, but it is the same money as the fee.
Funding. Perpetuals settle funding on a fixed interval, usually every eight hours. Hold for three days and that is nine payments in or out, none of which show up in your entry or exit price. On a position with a thin margin, funding is frequently the difference between profit and loss.
Impact. If your order is large relative to the resting liquidity, your own buying moves the price while you fill, and the quote you saw is not the average you get. Retail size rarely feels this. It is why a system that works at $10,000 can stop working at $500,000 with no change to the code.
Model only the first of the four and the backtest is measuring something other than the trade you would actually make.
What multiplies it
Those costs are per trade, so how often you trade decides how much they matter. A strategy that trades once a day pays the whole stack around 250 times a year. If your true round-trip cost is five basis points higher than assumed, that is more than twelve percent a year of drag that appears nowhere in the research. Most strategies don't have twelve percent of edge to spare.
So turnover is worth deciding early rather than discovering later. Two systems with the same signal quality and different holding periods can have very different cost profiles, and only one of them may clear its own costs.
The second multiplier is what you trade. On the most liquid instrument in the market, the spread can be about a basis point and the book absorbs size without much complaint. Move down the liquidity ranking and spreads widen by an order of magnitude, depth thins out, and impact shows up at sizes that would be invisible in a major. Funding gets more extreme in thin markets too.
So the same signal can be a viable business on the top handful of assets and impossible on the fortieth. When a backtest equal-weights a wide universe, it reports the average of those two things, and the untradeable half usually contributes the best-looking returns. That makes universe selection part of cost modelling rather than a separate decision.
Where cost models go wrong
Two mistakes show up again and again, and neither of them is about the size of a cost.
The first is the assumption that you can avoid the spread by never crossing it. Quote passively, collect the rebate, let the market come to you. What gets left out is adverse selection. Your order sits in the book and fills when someone else decides that trading at your price is attractive, and that decision is often better informed than yours. When the market moves your way, your order tends to get cancelled ahead of you. When it moves against you, you get filled.
Passive execution still has its place. What breaks is the line in the backtest that says maker fills simply happen. That is one line in research and a real operational problem in production, and a strategy whose profitability depends on it is making a claim about execution quality rather than about the signal. Only forward trading settles that claim.
The second mistake I have made myself. Costs arrive on their own schedules: fees at execution, spread at entry and exit, funding on the venue's interval rather than on your bar frequency. Flatten those schedules onto whatever timestamp your data happens to use and the arithmetic breaks with no error message, usually in your favour.
I ran a strategy for weeks that showed a spectacular return before the reason surfaced. An eight-hour funding rate had been forward-filled onto hourly bars, and the backtest charged it on every bar, so every payment was counted eight times. Corrected, the whole result inverted. The longer version is in the post-mortem, but the short lesson belongs here: getting the magnitudes right is only half of it. They have to be attached to the right clock.
Model the costs first
The usual order is to find a signal, evaluate it, then apply costs and see what survives. I would argue for the opposite.
Model the stack first, per instrument, at the size you actually intend to trade. What comes out is a floor: the minimum edge an idea has to produce, at a given turnover, to be worth running. That number tells you which strategy families are even available to you before you look at a single signal. High-turnover ideas need either a large edge or unusually good execution. Slower ideas survive on much less, and their real constraint moves somewhere else, usually to capacity or to how long the edge lasts.
It is not the interesting part of research. It is the part that decides which of the interesting parts are worth doing.
At Pearlixa we model costs at their real intervals and against real per-instrument spreads, because a cost model that flatters the strategy is no use to whoever runs it. If that approach is useful to you, early access is open at pearlixa.com/early-access.
*Nothing in this post is investment advice. Cryptocurrency trading involves substantial risk of loss. All figures are illustrative examples, not quotes from any specific venue.*