Crypto Backtesting: What Changes When the Market Never Closes

10 min readQuantParadox research

The engine is the same and the assumptions are not. Four things about crypto break a testing setup built for currency pairs, and three of them push results in the flattering direction.

The short answer

Backtesting a crypto strategy differs from forex in four concrete ways — perpetual funding costs, exchange-specific price series, the absence of sessions and weekends, and a universe where tokens genuinely disappear — and each one changes the result rather than the presentation.

What does funding cost do to a backtest?

Perpetual futures charge a funding payment between longs and shorts at regular intervals, and a strategy holding positions across those intervals pays or receives it — which a backtest ignoring funding silently omits from every held position.

The size is not trivial for anything held longer than a few hours. Funding is typically settled every eight hours, and during a strongly trending period the rate on the crowded side can be substantial. A strategy holding a long through a euphoric run is paying for the privilege on every settlement.

The direction of the error is systematic rather than random. Trend-following strategies tend to be positioned on the crowded side, which is the side that pays — so ignoring funding flatters exactly the strategies most likely to be affected.

Modelling it requires the historical funding rate series for the specific instrument and exchange, applied at each settlement the position was open across. It is bookkeeping rather than difficulty, and it is the single most commonly omitted cost in crypto backtesting.

Which exchange's prices are you testing on?

Crypto has no consolidated tape, so every exchange has its own book and its own price — and the differences are larger and more persistent than the equivalent variation between forex brokers.

The practical consequence is that a strategy tested on one exchange's history and traded on another is being estimated rather than measured. For wide-stop strategies the gap is noise; for anything tight it is a meaningful fraction of the edge.

Liquidity differs as much as price. A strategy that works on the deepest venue may be untradeable at size on a thinner one, and the backtest — which fills any order instantly at the printed price — will not mention it.

The discipline is to test on the exchange you will trade, model that venue's fee schedule including any maker-taker distinction, and treat results from an aggregated index price as indicative rather than executable.

Does 24/7 trading actually change anything?

It removes weekend gaps and the session structure that forex strategies often rely on, and it replaces them with a different kind of unevenness that is easy to miss.

The good news first: no weekend gap risk. A forex position held over Friday faces a Monday open that can be far from Friday's close with no opportunity to exit. Crypto trades through, so stops behave as modelled.

The complication is that continuous trading is not uniform trading. Volume and volatility still follow a strong daily and weekly shape driven by where the participants are awake, and a strategy assuming every hour is equivalent will find its fills and its spreads are not.

The other consequence is subtle and affects any calculation involving periods. Annualising a Sharpe from daily returns uses 252 for equities and 365 for crypto, and using the wrong constant misstates the figure by roughly 20% — a difference large enough to change how a result is read.

What about tokens that stop existing?

Crypto has the equity survivorship problem in a more severe form: tokens are delisted, chains are abandoned, and projects fail completely, so a universe built from today's listings excludes every one that did not make it.

Testing an altcoin strategy against the coins currently listed on a major exchange is testing against a set pre-filtered for survival — and unlike equities, where a failure often means an acquisition at a discount, crypto failures frequently go to effectively zero.

The bias is largest for exactly the strategies people most want to test. A rule that buys sharp declines looks superb when every decline in the sample eventually recovered, and the ones that did not recover are the tokens that were removed from the list you built the universe from.

The requirement is the same as for equities: a point-in-time universe including delisted assets with their real final outcomes. Without it, the result is an upper bound and should be reported as one.

Do the standard method rules still apply?

Every one of them, and more strictly rather than less: out-of-sample validation, counting variants, realistic costs and adequate sample size are unchanged by the asset class.

One thing that genuinely differs is the amount of history available. Most crypto instruments have a few years rather than decades, and much of that history sits inside one or two enormous directional regimes. A strategy validated only across a bull market has met one condition.

That shortage makes the trial count more dangerous rather than less. Fewer independent periods and the same number of parameter combinations means a higher chance that the best variant is fitted, which is the opposite of the intuition that a newer market is easier to find edges in.

The compensating advantage is frequency. Continuous trading and high volatility mean trades accumulate quickly, so a strategy's sample size can reach a usable level in months — provided the regime coverage is acknowledged for what it is.

Questions people actually ask

Should I backtest on spot or perpetual futures?

Test the instrument you intend to trade, because the cost structures are genuinely different. Perpetuals carry funding and offer leverage; spot carries neither and settles differently. A strategy validated on spot prices and traded on perpetuals is missing a recurring cost that applies to every position held across a settlement.

How much crypto history is enough?

Enough to include more than one regime, which for most instruments means covering at least one major decline as well as an advance. Calendar length matters less than regime coverage — three years spanning a full cycle is more informative than five that were uniformly upward, and a strategy that has only met rising markets has learned that buying works.

Do exchange fees matter as much as forex spread?

They matter in the same way and are easier to model, since published maker and taker fees are explicit rather than embedded in a quote. What is easy to miss is the maker-taker distinction: a strategy assuming limit fills earns the maker rate, and if it actually crosses the spread it pays the taker rate instead, which can be several times higher.

The only backtest that settles it is yours.

Build a strategy from a sentence, paste your own Python, or import your live trade history and have it graded. Five full backtests free, no card, and we'll tell you plainly when the result is indistinguishable from luck.

We publish research and tooling, not trading advice, and we make no claim about future returns. Everything above describes how to test an idea — not a reason to trade one.