Backtesting
The edge is usuallyreal. The costsare usuallyimaginary
Most failed strategies were not wrong about the market. They were wrong about what it costs to act on it — and every one of those costs is a number you can measure instead of assume.
The four costs, in the order they surprise people
Fees, which are not what the venue publishes
One venue we trade publishes a zero-fee schedule for a market where our own fills show about 14 basis points round trip. A backtest using the published number is optimistic by that amount on every trade, which is enough to invert a high-frequency edge entirely. The effective rate here comes from the fills, not the schedule.
Slippage, which scales with your size and the book
Filling at the mid is a fiction. A backtest should pay the spread and then some, and the "then some" depends on how much of the book you are taking. Paper here applies slippage and refuses a market order larger than the depth it is modelling, rather than filling it at the mark and flattering the result.
Capture — the gap between the signal price and the fill
The most under-modelled cost. Your rule fires at a price; your order reaches the venue later, and at a size that moves the book. Measured on our own execution, the majors round 13–32% off at our size, while several mid-caps come in clean. Capture, not fees, was the biggest profit thief we found.
Granularity, which is a per-symbol fact
Minimum sizes and tick rounding mean the order you can actually place is not the order the strategy computed. On a cheap asset that is noise; on an expensive one at a small account size it can be a third of the position. This is why the paper venue borrows the real instrument specifications: so the rehearsal rounds the way the real thing will.
And one that is not a cost
A test window that contained one regime
A mean-reversion rule tested through a range looks superb and dies in the first trend. Walk-forward exposes some of this; the rest is a data problem. With about two years of crypto history, the number of genuinely independent regimes available is small, and pretending otherwise is the failure that no amount of cost modelling fixes.
What we found when we scored our own signals
An audit of 857 recorded signals scored worse than random against the outcomes, and the model verdicts attached to them separated nothing. Publishing that is uncomfortable and it is the reason the rest of this page is specific: the discipline exists because the naive version was measured and it did not work.
Questions about backtest realism
What fee should I model?
Your own, from your own fills, per venue and market type. Failing that, the venue’s taker rate rather than its maker rate — assuming you get the maker fill is the second-commonest optimism after ignoring slippage.
How do I estimate capture without trading?
Paper, at the size you intend to trade. It is why the free tier includes the whole execution path: capture is not a number you can reason your way to.
Does the platform stop me deploying a strategy that ignores this?
It will not let it promote itself, and the risk gate will size it. It will not stop you from trading a bad idea deliberately — but paper is free and the costs above are applied there.
Re-run a strategy with the costs on
Free paper accounts pay fees, take slippage, and refuse what the book cannot absorb.