Statistical Arbitrage in Crypto: What Survives After Costs

The Short Answer

Statistical arbitrage is not arbitrage in the risk-free sense. It is a probabilistic bet that a historical price relationship between two assets will reassert itself. In crypto it is almost always executed with perpetual futures on the short leg, which means funding payments are part of your cost structure - not a footnote. Whether a trade is viable depends less on the entry signal than on how fast the spread reverts relative to what you pay to hold it.

Statistical Arbitrage vs Spot Arbitrage

These get lumped together as "arbitrage" and they are fundamentally different trades.

	Spot arbitrage	Statistical arbitrage
What you observe	A price difference existing right now on two venues	A spread that is wide relative to its own history
Nature of the edge	Execution: can you complete the round trip	Probability: will the relationship hold
Can it simply fail?	The gap closes and you make nothing	The spread keeps widening and you lose
Hold period	Minutes to hours	Hours to weeks
What kills it	Fees, transfers, thin books	Costs accumulating while reversion does not arrive

Spot arbitrage is closer to a logistics problem - we cover its mechanics in the spot arbitrage guide. Statistical arbitrage is a forecasting problem wearing arbitrage clothing. Treating the second like the first is how people get surprised.

What Is Actually Being Traded

Statistical arbitrage compares two different coins, not the same coin in two places. That separates it from the other arbitrage types on this blog:

Strategy	Example legs	What you bet on
Spot arbitrage	BTC on Exchange A vs BTC on Exchange B	Same asset, two prices, gap closes
Funding / basis arbitrage	BTC spot vs BTC perpetual	Same asset, funding payments and basis
Statistical arbitrage	ETH vs SOL (or any two related coins)	Two different assets, their relationship returns to normal

A typical setup: ETH and SOL usually move together. Over the last few weeks SOL has run up much faster than ETH, more than their history suggests it should. A pairs trader goes long ETH and short SOL, betting the gap between them shrinks back. If both coins then rise or fall together, the two legs mostly cancel out. You only make or lose money on the relative move.

The example pair is for illustration only, not a trade recommendation. The rest of this article uses "asset A" and "asset B" for the two legs.

The Four Pieces of a Pairs Trade

1. Pair selection: correlation is not enough

The first filter is whether the two coins have historically moved together. You measure that with the correlation of their returns, where each return is the percentage price change from one period to the next (the log version below is the standard form):

Here $P_t$ is the price of one coin (say ETH) at time t, and $P_{t-1}$ is its price one period earlier. Compute this for both coins, then measure how closely the two return series line up.

High correlation means the two assets tended to move together. It does not mean their spread returns to an average. For that you want cointegration: the idea that some linear combination of the two price series is stationary, so deviations get pulled back.

The practical difference matters. Two assets can be 0.9 correlated while their ratio drifts steadily in one direction for months. Trading that as a mean-reversion pair means financing a position that never converges.

2. Hedge ratio: how much of each leg

You are not trading equal dollar amounts because the two assets do not move one-for-one. If SOL typically moves 1.5% for every 1% ETH moves, equal dollar legs leave you net exposed to SOL. Regressing one log price against the other gives that ratio (β, the hedge ratio), and what is left over is your spread:

If the hedge ratio is wrong, your "market-neutral" position carries directional exposure you did not intend - which is how a pairs trade turns into an accidental bet on the market.

3. Entry signal: the z-score

Standardise the spread against its own recent behaviour:

Convention is to enter when $|z|>2$: the spread is unusually wide, so you go long the underperformer and short the outperformer. This is the part every guide covers, and it is the least important part of the decision.

4. Exit rule: reversion, or a time stop

You exit when the spread returns toward its mean, or when a predetermined time limit expires. The time stop is not optional. Without it, a structural break - one asset repricing permanently - becomes an open-ended loss.

Why Crypto Changes the Math

Three crypto-specific facts reshape textbook stat arb:

1. The short leg is usually a perpetual future. You cannot easily borrow and short most altcoins on spot. So you short a perp - and inherit funding payments for the entire hold. Funding is a recurring cost or income depending on sign, settling every 8 hours on most venues. If you do not know what that means, start with funding rates explained.

2. Correlation regimes collapse under stress. In a sharp drawdown, most crypto assets correlate toward 1 against BTC. Precisely when your spread widens most dramatically, the statistical relationship you relied on is least reliable.

3. Markets run 24/7 with no closing auction. There is no daily reset, and liquidity varies enormously by hour. A spread measured on thin overnight liquidity can be an artefact rather than an opportunity.

The Cost Stack Nobody Nets Out

A pairs trade has two legs opened and two legs closed - four fee events - plus funding for the whole hold.

Illustrative arithmetic on a $5,000-per-leg position, 0.05% taker fees, funding running against you at 0.01% per 8-hour interval:

Target spread capture (1.5% of one leg):    +$75.00
Entry fees (2 legs x 0.05%):                 -$5.00
Exit fees (2 legs x 0.05%):                  -$5.00
Funding on short leg, 4 days (12 intervals): -$6.00
------------------------------------------------------
Net if the spread fully converges:          +$59.00
Net ROI on $10,000 committed:                  0.59%

Now change one assumption. If convergence takes three weeks instead of four days, funding alone costs roughly $31.50, and a partial convergence of half the spread leaves you at about break-even. The spread did most of what you predicted, and you still made nothing.

That is the honest shape of stat arb: the entry signal was right, and the cost of waiting consumed the edge.

Half-Life Is the Number That Decides It

If reversion speed determines profitability, you should measure it rather than hope. Fit the spread to a mean-reverting process and derive its half-life - the time for roughly half of a deviation to decay:

Then compare it against your cost per unit time:

• Short half-life (hours): costs stay small relative to the expected capture. Tradeable.
• Medium half-life (a few days): viable only if the spread is genuinely wide and funding is neutral or favourable.
• Long or unstable half-life: not a mean-reversion trade at all. You are holding a directional position and paying for the privilege.

We use the same measure - alongside z-score, index correlation and maximum adverse excursion - to score cross-exchange opportunities. The full set, with formulas, is in cross-exchange arbitrage metrics.

Where Retail Stat Arb Actually Breaks

Overfitting during pair selection. Test 500 pairs and dozens will look cointegrated by chance. Unless you hold out data and validate on a period you never looked at, you have found a coincidence, not a relationship.

Survivorship in the candidate list. Backtesting today's listed assets ignores everything that was delisted. Delistings are not random - they are exactly the structural breaks that destroy pairs trades.

Costs modelled as an afterthought. Published stat-arb results in crypto sometimes quote returns of 10%+ per month. Net out four fee events, funding on the short leg over realistic hold times and slippage on the thinner leg, and those figures shrink dramatically. Any result that does not itemise those four costs is not a result.

Capital inefficiency. Two legs plus margin buffer means a lot of committed capital per position - and returns should be calculated on the whole committed amount, not on one leg.

Discipline under drawdown. The trade is designed to move against you before it works. Without a pre-committed size and time stop, "it will revert" becomes the most expensive sentence in your trading log.

Is It Worth It Compared to Simpler Strategies?

Honestly: for most people, not first.

• Spot arbitrage has a smaller, more certain edge and its main challenge is execution rather than forecasting.
• Funding rate arbitrage earns a scheduled cash flow, and its risks are measurable before entry - see how to evaluate a funding trade.
• Statistical arbitrage adds a forecast on top of the same operational burden as both.

If you have not yet worked through which of the simpler two fits your capital and attention, spot vs funding arbitrage is the more useful starting point. Stat arb is a reasonable next step once you are already netting out costs reliably and reading realistic profit expectations, which we cover in is crypto arbitrage profitable.

Frequently Asked Questions

Is statistical arbitrage the same as crypto arbitrage?

No. Spot arbitrage captures a price difference that exists right now on two venues, so the edge is observable before you commit. Statistical arbitrage bets that a historical relationship between two assets will reassert itself, which may or may not happen. One is an execution problem, the other is a probability estimate with real loss scenarios.

Do I need cointegration or is correlation enough?

Correlation alone is not enough. Correlation says two assets tended to move in the same direction; cointegration says their spread has a stable long-run equilibrium it returns to. Trading on correlation without cointegration means you can be long a spread that simply keeps widening, because nothing structurally pulls it back.

How does funding affect a crypto pairs trade?

In crypto the short leg is normally a perpetual future, so you pay or receive funding for the entire holding period. On an 8-hour interval that is three payments per day against your position. If your expected convergence takes four days and funding runs against you, the funding cost alone can exceed the spread you were trying to capture. The funding rate calculator is the fastest way to size that cost before entering.

What z-score should I use to enter?

A z-score above 2 is the common convention for "unusually wide", but the threshold matters less than pairing it with a reversion estimate. A 2-sigma spread with a short half-life is a trade; the same 2-sigma spread with a long or unstable half-life is just a wide spread that may stay wide while your costs accumulate.

Is statistical arbitrage realistic for retail traders?

It is possible, but it is the least forgiving member of the arbitrage family. You need capital for two legs plus margin, clean historical data, the discipline to size and stop out, and the honesty to reject pairs that only worked in a backtest. Most retail traders get better risk-adjusted results from simpler spot or funding strategies first.

Why do backtested stat-arb returns rarely survive live trading?

Three reasons dominate: transaction costs and funding are understated or ignored; pairs are selected by searching many combinations until something fits, which is overfitting; and correlation regimes shift, so a relationship that held for two years breaks during a month of market stress.

Summary

• Statistical arbitrage is a probabilistic strategy, not a risk-free one. The spread can widen indefinitely.
• Correlation identifies candidates; cointegration is what justifies a mean-reversion trade.
• The z-score entry is the easy part. Reversion speed versus holding cost is the decision.
• In crypto the short leg is a perpetual, so funding is a core cost, not a detail - three settlements a day on most venues.
• Four fee events plus funding plus slippage is the real cost stack. Any published return that does not itemise them should be treated as marketing.
• If you are choosing your first market-neutral strategy, simpler spot or funding carry usually offers better risk-adjusted returns for the same effort.