An honest post-mortem of a market-neutral crypto statarb hunt, and the three fake edges I almost believed.
In about a month I built a market-neutral statistical-arbitrage engine for liquid crypto mid-caps and tested every classic strategy family against it: reversal, momentum, funding carry, funding momentum, cointegration pairs, two flavors of lead-lag, and an adaptive multi-signal composite. Eight strategies. All eight failed once costs, funding, and fills were modeled honestly.
That sounds like a story about failure. It isn't. Along the way the harness caught three fake edges before a single dollar of real money touched them: strategies showing +604%, Sharpe 76, and "profitable every single year". Each one died to a specific, repeatable rigor check. Those checks are the actual product of this project, and they transfer to any backtest you will ever run.
One thing to know before the numbers. This project was deliberately scoped to what a solo trader or a small team can realistically build: standard methods, executed carefully, with honest accounting. No exotic math, no alternative data, no colocated execution. That is the toolkit most people testing strategies at home actually have, so that is the toolkit I held to.
Here is the whole hunt, strategy by strategy, with the real numbers.
The setup
The goal was the classic uncorrelated sleeve: a dollar-neutral, beta-neutral statistical-arbitrage book on crypto mid-caps, holding positions from a few hours to about a day.
The testbed every strategy had to pass:
- Data: three years of hourly bars on roughly 100 liquid perpetuals, with real funding
- Universe: a dynamic daily liquidity screen. Coins below a dollar-volume floor drop out at
- Costs: maker fee plus half the actual per-coin spread, funding charged at its real
- Validation: strict walk-forward with no look-ahead, quarterly breakdowns, permutation null
The engine itself is real and reusable, with over a hundred unit tests behind it. The machine wasn't the problem. The market was.
The eight strategies
1. Cross-sectional reversal
The bread-and-butter statarb trade: fade the moves. Buy the coins that fell against the market, short the ones that popped.
It initially looked alive: +24.9% per year net at maker fees. Then the walk-forward split told the real story. One single quarter carried the entire year, two of five quarters were negative including the most recent one, and extending to the full sample showed 2023 and 2024 both negative. A regime filter didn't rescue it; the failures weren't regime-detectable.
Verdict: one lucky quarter, not an edge.
2. Cross-sectional momentum
The mirror trade: ride the moves instead of fading them. If reversal is noisy, maybe the tail trends.
It doesn't. Negative net over the full sample. Fastest kill of the eight.
Verdict: dead on arrival.
3. Funding carry, the +604% that wasn't
Short the coins with the highest funding rates, collect the funding they pay, hedge dollar-neutral with the low-funding side. This one is the heart of the story.
It looked spectacular. Over the full sample, with the liquidity screen and real per-coin spreads: +604%, Sharpe 3.73, max drawdown of 11%, positive in all but one quarter. It survived the 2024 memecoin alt-season. It was crash-protective in the 2025 downturn. I stress-tested it, decomposed it, and built the live execution for it.
Then one number refused to reconcile, and the bug surfaced. Perp funding settles every 8 hours, but the data pipeline forward-filled the 8-hour rate onto every hourly bar, and the backtest booked it on every bar. Every funding payment was counted eight times. The entire edge was the multiplier.
Corrected, the same strategy shows −51%, Sharpe −1.26, a 56% max drawdown, and only a handful of quarters positive. The decomposition after the fix explains why the trade can't work: shorting coins that are pumping hard enough to have extreme funding costs about 68% per year in price P&L, while the real funding collected is only about 46% per year. You pay more in price than you earn in carry.
Verdict: the most convincing backtest of the project was an accounting bug.
4. Funding momentum
The inverse trade: go long the high-funding coins, treating extreme funding as a crowding signal, and ride the enthusiasm instead of fading it. After the bug fix inverted carry's sign, this direction briefly looked hopeful on paper, with the price leg earning more than the funding leg cost.
The full engine test killed it: negative, with real costs, real spreads, and walk-forward. The price edge is not stable enough to reliably out-earn the funding you now pay to hold the position.
Verdict: both directions of the funding trade lose. The funding market is priced correctly.
5. Cointegration pairs
The textbook classic: find cointegrated pairs in-sample, trade the spread mean-reversion out-of-sample. Done rigorously, with pairs selected only on the in-sample window and traded only on data they had never seen.
Result: −16% out-of-sample before costs, −55% after. Not "profitable but eaten by fees". Negative gross. Crypto cointegration relationships simply don't persist out of sample. The pairs that look locked together for six months break the moment you rely on them.
Verdict: the relationship itself is unstable. No cost model can fix that.
6. Price lead-lag, majors to tail
Do BTC and ETH moves propagate into the mid-cap tail with a tradeable delay? Buy the laggards when the majors move?
Positive in exactly one of three years. That's not a lag structure. That's a coin flip with extra steps.
Verdict: fluke.
7. Funding lead-lag
Does the funding level predict forward price? Signal on funding, trade the price. This looked genuinely interesting at first: +36% per year, measured on price P&L only.
Then I charged the funding you actually pay while holding those positions: −44% per year. The "signal" was just the carry trade wearing a costume. The positions it puts you in are long high-funding coins, and the funding bill erases the price edge. I named this failure mode *the carry illusion*, because it showed up twice.
Verdict: a funding signal that ignores funding costs isn't a signal.
8. The adaptive multi-signal composite
The last resort of every quant: if no single signal works, combine twelve of them. Reversal, momentum, funding, volume, and volatility signals, weighted by recent predictive quality, adapting as regimes shift.
It produced the best-looking equity curve of the entire project: +198%, Sharpe 2.72, positive in every year of the sample. Every sub-signal contributing, weights adapting sensibly. This was the one that felt like the payoff for months of infrastructure.
Then I ran the implementation-lag test: re-run the exact same backtest, but enter each position one bar after the signal fires. That is the honest simulation of a simple fact: you cannot trade the close price of the bar that generated your signal.
+198% became −180%. Every winning year became zero.
The "edge" was bid-ask bounce. Signals fire when the close prints on the wrong side of the spread, and the backtest "captures" the bounce back to mid, a fill no real order would ever get. Restricting to only the slow signals didn't help: −179%. The entire composite was harvesting an artifact of close-price fills.
Verdict: the best backtest of the project couldn't survive a one-bar delay.
The side quest: a Sharpe 76 basis trade
One more deserves its own confession, because it produced the single most absurd number of the project. A quick test of the perp-versus-spot basis showed a "premium" you could apparently harvest with Sharpe 76.
Nobody has Sharpe 76. The number itself is the bug report. The "premium" was the two legs printing on opposite sides of their spreads, so the backtest was buying at one market's bid and selling at the other's ask in the same instant. That trade exists on paper and nowhere else. The same family of idealized-fill artifacts came back later in another test: an edge that scored around Sharpe 28 with idealized passive fills scored −6 once fills were modeled realistically.
Verdict: too-clean numbers are guilty until proven innocent.
The three fake edges, and the checks that caught them
| Fake edge | Headline number | What it actually was | The check that killed it |
|---|---|---|---|
| Funding carry | +604%, Sharpe 3.7, all but one quarter positive | 8-hour funding booked on every hourly bar, counted eight times | Verify every funding magnitude against the real settlement interval |
| Basis trade | Sharpe 76 | Two legs printing on opposite sides of the spread | The lag test, plus the too-clean-number rule |
| 12-signal composite | +198%, positive every year | Bid-ask bounce harvested via close-price fills | The implementation-lag test: enter one bar later |
Three different strategies, three different bugs, one shared property: each one was the most exciting result in the project at the moment it appeared. That is not a coincidence. Your best-looking backtest is the one most likely to contain an error, precisely because errors are unbounded and real edges are not.
The rigor checklist
Earned the hard way. Skip none of these:
- •Verify every funding and cost magnitude at its real interval. Never forward-fill a rate
- •Run the implementation-lag test. Re-run entering one bar after the signal. If the edge
- •Net of real fills and real spreads, always. And put a liquidity floor under the universe.
- •A Sharpe above 3, or any too-clean number, is guilty until proven innocent. Sharpe 76 is
- •No fishing. Testing variant after variant until one shows green is the fake-edge
- •Permutation-null everything. If shuffled data produces your Sharpe, you have nothing.
- •Backtested fills are a claim, not a fact. A maker-fee cost model assumes your passive
What "no edge" is worth
The comprehensive verdict, after eight strategies over three years with honest costs: liquid crypto mid-caps are efficient enough that simple market-neutral statarb does not pay with the toolkit a solo trader can realistically build. Reversal, momentum, carry, funding momentum, pairs, both lead-lags, and the composite. None survives correct funding accounting plus realistic fills. I stopped generating variants at that point deliberately, because multiple-testing discipline cuts both ways.
That conclusion cost a month of focused work, and it was worth it, for three reasons.
First, the negative is real information. Knowing that this entire strategy class is priced out for a small player, from my own harness with my own costs rather than from a paper, means never again burning capital or attention on its variants.
Second, the harness is the asset. A pipeline that catches a +604% accounting bug, a Sharpe-76 artifact, and a 4-of-4-years mirage before live money is a working fake-edge detector. Every future idea inherits it, and the marginal cost of testing the next idea honestly is now near zero.
Third, it recalibrates where to look. The edge a small player can win with is not better signals on the most liquid instruments. That axis is picked clean, and everyone's backtest says otherwise right up until the lag test.
This is also, one level up, why Pearlixa exists. The gap between a retail backtest and institutional validation isn't intelligence. It's infrastructure and discipline: funding checked at its real interval, the lag test run every time, the too-clean number treated as an accusation. Pearlixa's validation runs on that same discipline, and models that fail it are models you never see. If you want the output of that process without spending months building the harness yourself, early access is open at pearlixa.com/early-access.
*Nothing here is investment advice. It's the opposite: evidence that the strategies everyone backtests first don't survive contact with honest accounting.*