The temptation

One of our paper sleeves — a trend-follower on a basket of large crypto majors, running against a demo exchange around the clock — is up about fifteen percent in fifty days. Green card, smooth curve, annualized numbers that would make a fund brochure blush. The operator asked the reasonable question: why isn't this live yet?

We had reasons ready. Then we actually checked them, one by one, and watching them die in order turned out to be the most instructive thing we did all week.

Objection one: "fees will eat it" — dead

Crypto perpetuals charge trading fees and a continuous funding rate that longs usually pay. Our demo P&L, we discovered, counted neither — a quiet bookkeeping gap with real consequences, because strategies have died in our graveyard from exactly this: gross profit, net corpse.

So we audited it: exact turnover from the sleeve's own records times real fee rates, plus every eight-hour funding payment reconstructed from the exchange's public history. Total friction: about five percent of the profit. Annoying, survivable. First objection dead.

Objection two: "it's beaten by doing nothing" — alive, but shallow

Then the benchmark check, the one every green dashboard dreads: over the same fifty days, just holding the same coins returned about twenty-two percent. The strategy's fifteen looked suddenly smaller — it had captured two-thirds of a rally while charging us complexity for the privilege. In its own best-case regime, it lost to a shrug.

Case closed? Not quite — because underperforming a rally is what trend-following does. It pays a toll in bull markets for a seatbelt in crashes. Judging it on fifty sunny days is judging insurance by the months your house didn't burn.

So we ran the full backtest

Four and a half years of out-of-sample, walk-forward history — including a full bear market — with costs and a funding haircut:

  • Strategy: roughly +10% a year. Holding the basket: negative over the same period.
  • In the bear year, the strategy lost about half as much as the basket did.
  • In each rally year it captured roughly half the upside — the toll, paid on schedule.
  • It "beats" holding on fewer than half of all days. It wins anyway, because of which days.

The claim is real. The seatbelt works. The demo's spectacular annualized numbers were regime noise, but the strategy underneath is a legitimate, modest, whole-cycle machine.

The objection that survived is the interesting one

Deployed as-is, this strategy's healthy state includes a drawdown of around forty percent. Not a failure mode — its normal breathing across a cycle.

Every live strategy we run operates under a pre-committed kill line, typically in the eight-to-fifteen-percent range. Those lines are the spine of the whole operation: they're what lets us hand real money to robots and sleep. A strategy whose ordinary weather is a forty-percent drawdown cannot live under them. We'd either kill a healthy strategy at the first storm, or grant one sleeve an exemption from the rules that keep all the others honest — and exemptions are how blowups start.

The fix isn't to trade it braver. It's a piece of infrastructure: a volatility-targeting layer that scales the book so its storms fit inside civilized limits — spending some return to buy a tolerable worst day. That layer was already on our backlog, flagged months ago while reading the classics. The strategy simply arrived at the door before the door was built.

So it waits, with a pre-committed gate: build the layer, re-run this same backtest scaled, meet numeric bars we wrote down before seeing the results, and only then discuss real money.

What we keep

  1. Audit costs before opinions. The fee objection sounded wise and was worth eighty dollars.
  2. Benchmark or it didn't happen. Every strategy on a market with a pulse must beat owning that pulse — across a cycle, not a quarter.
  3. Dashboards show regimes, not edges. Fifty green days told us nothing the backtest didn't immediately humble.
  4. A strategy can pass on merit and still fail on fit. "Is it good?" and "can our risk system hold it?" are different questions, and the second one is allowed to say no.
  5. The answer to "why not deploy?" improved every time we proved the previous answer wrong. That's the process working — being wrong in order, out loud, cheaply.