Take your best backtest and flip the sign on every position. Longs become shorts, shorts become longs. Run it again.
If the flipped version loses roughly what the real one made, good. If it still makes money, you have a problem.
What this test is actually asking
A trend signal claims to know something about direction. That's the whole claim. So if you strip out the direction and the result survives, the signal wasn't providing it.
The cleaner version is to randomize. Keep every entry time, every exit time and every position size exactly as they were. Only replace the long-or-short decision with a coin flip. Do that a thousand times and you get a cloud of placebo equity curves.
Now ask one question. Where does your real curve sit inside that cloud?
Why random can look so good
Crypto rewards a lot of things that aren't prediction. The market drifts, it clusters volatility, and it has long stretches where almost everything moves together. A strategy that is simply invested in the right periods and sized well can look clever without ever picking a direction well.
Here's a made-up but typical example. Say your real backtest returns 180% over four years. You run 1,000 random-direction versions and the median comes back at 95%, with the best ones above 200%. Your real result isn't special. It's sitting in the middle of noise that happens to be positive.
That's a volatility bet. You found a way to be active when the market was active. Fine, but that's not a trend edge, and it won't behave like one when conditions change.
The mistake I see most
People compare their strategy to zero. Did it make money? Yes? Ship it.
Zero is the wrong benchmark. The right one is what a blind version of the same system would have earned. A long-biased market hands out free money to anything that leans long at the right times, and a flat benchmark never shows you that.
I'd also be careful with a single placebo run. One random draw can land anywhere. You need the distribution, not one lucky or unlucky sample.
How to read the result
Look at your real result against the placebo curves, and check two things. First, is it above, say, the 95th percentile of the random runs? Second, is the gap big enough to matter after costs?
If the real curve is near the pack, the signal isn't doing much. Don't reach for more parameters to fix it. That just overfits the noise you were trying to detect.
A real trend system should show a clear gap on the Sharpe ratio too, which is return divided by volatility, a rough measure of how much you earned per unit of bumpiness. If random versions match your Sharpe, you're being paid for the ride, not the call.
One more thing. Run the placebo on the same data you used to pick your parameters, then again on data you never touched. A signal that beats the cloud in-sample and falls back into it out-of-sample was never telling you anything. I've thrown away more than one favorite this way, and it stung every time.
Do this before you go live
Pick your best candidate. Run the flipped version first, since it takes five minutes. Then build the randomized cloud with at least a few hundred draws.
If you fail it, that's useful news, and cheap news. It's far better to learn it from a script than from a drawdown with real money in it.
It's one of the checks I run on everything before it goes near a live account. If you want to see how we think about validating results, the published methodology is at v33systematic.com.