Is Crypto Arbitrage Still Profitable in 2026? I Tested It (And 5 Other Strategies) With Real Data

A physicist backtested every popular crypto trading strategy out-of-sample on real exchange data. Here’s what actually survived — and why your backtest is probably lying to you.

Every crypto strategy looks profitable until you trade it with real money.

You’ve seen the thumbnails: the “$500/day arbitrage bot,” the RSI strategy with the 70% win rate, the backtest equity curve that only goes up and to the right. Then you deploy real capital and it bleeds out within weeks.

I have a PhD in physics, and I got tired of not knowing whether any of these strategies actually work. So I built proper backtests for the most popular ones — on real exchange data from Coinbase, Kraken, Kraken Futures, and Deribit — and then I ran the one test almost no strategy seller ever does.

This article walks through what I found. The short version: almost every popular strategy is a statistical illusion, and once you understand why, you’ll never be fooled by a backtest again.

What “out-of-sample” means — and why it’s everything

Here’s the trick that makes fake strategies look real.

Take any strategy with adjustable parameters — a moving-average length, an RSI threshold, a lookback window. Run it over historical data, and tweak the parameters until the backtest looks great. With enough combinations, something will always look profitable. That’s not skill. That’s the optimizer finding random patterns in noise.

This is called curve-fitting (or overfitting), and it’s why your backtest lies. The strategy isn’t capturing a real market edge — it’s memorizing the specific wiggles of the historical data you tested on. When you deploy it on new data, the wiggles are different, and the “edge” evaporates.

The honest test is simple: optimize your strategy on the first 60% of history, lock the parameters, and then run it — untouched — on the final 40% the strategy has never seen. This is called out-of-sample (OOS) testing. If the edge is real, it shows up in both halves. If it was curve-fit, the OOS half exposes it instantly.

Almost no YouTube strategy, paid course, or Twitter thread ever shows you the out-of-sample result. Now you know why.

Strategy 1: Cross-exchange arbitrage — dead

The classic idea: Bitcoin is slightly cheaper on Kraken than Coinbase, so you buy on one and sell on the other, pocketing the difference. It sounds risk-free.

The reality: spreads between major exchanges are around 0.01–0.05%. Your round-trip trading cost (fees plus slippage) is around 1%. The spread is roughly 80 times too small to clear your costs. Over 90 days of hourly data, the number of profitable trades was zero.

This strategy died years ago, the moment professional market makers wired the major exchanges into millisecond-level lockstep. By the time you see a gap, a bot has already closed it.

Strategy 2: Triangular arbitrage — dead

The more sophisticated cousin: exploit a price inconsistency within one exchange across three pairs — say USD → BTC → ETH → USD. If the implied cross-rate disagrees with the direct rate by more than your fees, you profit.

The reality: across three legs, your break-even cost is about 2.10%. The maximum mispricing I observed in the data was 0.52% — and that was a rare spike, not a norm. Zero trades cleared costs at any timeframe tested. The triangle stays balanced because bots keep it balanced, continuously, faster than any human or retail script.

Strategy 3: Intraday bots (RSI, mean reversion, momentum) — curve-fit

This is the category that fools the most people, because it looks so convincing. I tested three classic intraday strategies on hourly Bitcoin, using near-zero maker fees (the best case any retail trader can get).

In-sample, they looked fantastic. Out-of-sample, they fell apart:

• RSI reversal — In-sample Sharpe +1.06 (69% win rate). Out-of-sample Sharpe −0.98.

• Mean reversion — In-sample Sharpe +0.72. Out-of-sample Sharpe −1.46.

• Momentum breakout — In-sample Sharpe +0.64. Out-of-sample Sharpe −1.15.

The RSI strategy showed a 69% win rate and +29.6% annualized in-sample. If someone showed you that backtest, you’d believe them.

Out-of-sample, all three flipped to negative expectancy. Same code, same near-zero fees — the only difference was data the strategy hadn’t been tuned on. This proves the in-sample “edge” was pure curve-fitting. And note: because fees were near zero, this isn’t a cost problem. The strategies simply had no real predictive power.

This is exactly how retail algo traders blow up: they backtest, see the green numbers, deploy, and slowly bleed out as reality refuses to match the fitted past.

Strategy 4: Cross-sectional momentum — even the academic edge failed

Surely the strategies with peer-reviewed academic backing work? Cross-sectional momentum — ranking a universe of coins by recent return and buying the winners — is documented in the Journal of Finance (Liu, Tsyvinski & Wu, 2022) as a genuine crypto factor.

I tested it on 26 liquid, actually-tradeable coins:

• Long-only momentum — In-sample +54%/yr (Sharpe 0.93). Out-of-sample −77%/yr (Sharpe −1.65).

• Long-short momentum — In-sample +37%/yr (Sharpe 0.91). Out-of-sample −58%/yr (Sharpe −1.96).

Even the academically-documented edge collapsed out-of-sample. The catch with the published research: the effect concentrates in tiny microcap coins you can’t actually trade at any size, and in market regimes that have since reversed. On a tradeable universe of liquid coins, it doesn’t survive.

Strategy 5: Dominance rotation — fee-murdered

The popular “alt season” play: rotate between Bitcoin and a basket of altcoins based on the ETH/BTC trend. When alts are outperforming, hold alts; when Bitcoin leads, rotate back.

The reality over the test window: altcoins catastrophically underperformed Bitcoin (ETH −25%, AVAX −86%, LINK −70%, versus BTC +96%), and the strategy’s 204 rotations stacked up roughly 510% in cumulative fees. The result was a 99% loss of capital. Signal-based rotation is brutally regime-dependent and gets murdered by transaction costs.

The two strategies that actually survive

After everything failed, exactly two approaches held up out-of-sample. They share one defining trait: they don’t predict price direction. They get paid for providing a service.

1. Funding-rate carry (delta-neutral). Hold spot, short the perpetual future of the same coin, and collect the funding rate. Price movement cancels out — it’s market-neutral. This survives precisely because there’s nothing to curve-fit: you’re paid to hold a balanced book, not to forecast anything. Realistic net yield is a modest ~3–5% per year on majors, and it’s regime-dependent (you sit out when funding is thin or negative).

2. Covered-call premium (when volatility is high). Hold Bitcoin, sell out-of-the-money call options, collect the premium. Real income — but only worth it when implied volatility is elevated. The best time to sell is right after a crash, when fear makes premiums richest. In calm markets the premium is too thin to bother.

Notice what both have in common with how professional trading firms make money — market making, basis trades, liquidity provision. None of them predict direction. They all earn a spread, a premium, or a fee for providing capital and bearing risk. That is the actual secret the losing majority never learns.

How to never get fooled by a backtest again

Three rules. Internalize them and you’ll save yourself more money than any single strategy could make you:

1. In-sample profit is worthless. Demand an out-of-sample result. No OOS test = marketing, not evidence.

2. Model every cost explicitly. Fees, slippage, and funding. A strategy trading 200 times a year at retail fees carries an ~87% annual cost drag — dead before it starts.

3. Distrust anything that predicts price direction. Directional prediction is the most efficient, most competed-over part of any market. The edges that survive get paid for a service instead.

The honest bottom line

There is no retail strategy that turns a small account into a money printer. The directional strategies don’t survive out-of-sample. The market-neutral ones are real but modest, and scale only with capital. The genuinely large edges — market making, HFT — require infrastructure, not insight.

But the most valuable thing isn’t a strategy. It’s a filter — the ability to tell a real edge from a curve-fit fantasy. That filter is worth more than any signal.

I packaged the full report — every result, every number, the mechanism behind why fake edges fail — together with the complete Python toolkit I used to produce all of it (six backtesters plus a live market-neutral scanner, all running on free public APIs). If you want to verify these results yourself or stop being sold curve-fit fantasies, it’s here: https://ncappelluti.gumroad.com/l/honest-crypto

Educational only. Not financial advice. Past performance — especially in-sample past performance — does not predict future results. That is, in fact, the entire point.