Our GPU found a +172% strategy — and it was a lie

The memory of a champion

Months ago, a GPU-powered genetic algorithm produced what we remembered as "our first highly profitable CUDA strategies". The memory was precise: spectacular equity curves, a champion far above the market. We decided to verify that memory. Good thing we did.

The reproduction: 24 million backtests in 3.5 seconds

On an RTX 4080, we replayed the search: 24 million backtests in 3.5 seconds, a genetic algorithm crossing, mutating and selecting strategy variants on real Bitcoin data.

The champion came back, faithful to the memory: +172.7% on its training period. The kind of number that sells courses.

Then we did the only thing that matters: we released it on data it had never seen.

Out-of-sample result: −37%. Worse than doing nothing. The champion wasn't a strategy — it was a memory dressed up as a strategy.

The autopsy: where the lie was manufactured

The bug isn't in the GPU. It's in the selection function. Here, untouched, is the original engine's fitness formula:

fitness = total PnL × 0.4
        + win rate × 10 × 0.3
        + (100 − max drawdown) × 0.2
        + trade count × 0.1

Four terms, and not once the question: "on which data?" Selection scored every variant on the same candles it was optimizing. No split, no holdout, no sealed data.

With 24 million attempts on a fixed dataset, you are no longer searching for a strategy: you are searching for the parameter combination that best memorizes the past. And you always find it. That's a mathematical property, not bad luck: the more variants you test on the same data, the further the best observed score drifts from true expectancy. The +172% didn't measure an edge — it measured the size of our search.

The test that unmasks any mirage

One protocol separates a strategy from a memory:

1. Seal a slice of data BEFORE the search begins.
2. Search, mutate, optimize as much as you want — without ever touching it.
3. Open the seal once, at the end, on the champion.
4. Accept the verdict, even if it ruins weeks of work.

Step 4 is the hard one. It's also the one 99% of published backtests skip — your platform's Strategy Tester doesn't force you into it: it's the author's tool, not a judge.

Put your own strategy through this test

• Free Windows Lab: reproduce this protocol on your machine — your GPU explores the variants, a sealed holdout stays invisible during the search, and the CPU delivers the local verdict. Educational simulation, no profit promises.
• Free Quick Check: paste a Pine script, no account. For an independent server-bound execution with a sealed holdout and a private immutable report, the audit exists — and if your strategy dies on the holdout, that's a delivered result, not a service failure. Nobody can buy a better verdict.

The +172% makes better advertising. The −37% makes better decisions.