Published
Here's a number that sounds impressive until you understand it: I recently tested about 85 million trading strategies in a single afternoon. One NVIDIA RTX 3070 — a gaming card — chewing through 386 crypto pairs, 220,000 configurations each, 1,600 simulated entries per run, in roughly 4 hours and 51 minutes. Do the multiplication: 386 × 220,000 is close to 85 million little experiments.
You'd think that much compute would surface something special. It mostly surfaced a lesson I keep relearning: more searching doesn't find an edge, it manufactures mirages.
Why brute force lies to you
When you test 85 million combinations, a few thousand will look incredible by pure chance. Flip enough coins and some will land heads ten times in a row. The strategies that bubble to the top of a giant search aren't the ones with real predictive power — they're the ones that best fit the random noise of the exact slice of history you tested on. Hand that "winner" the next month of data and it falls apart. This is overfitting, and the bigger your search, the worse it gets. Scale is not a substitute for an edge; it's a faster way to fool yourself.
What actually separated signal from noise
Three filters did almost all the work of telling junk from substance, and none of them was the win rate:
1. Net after fees. Most "profitable" configs were profitable in a frictionless dream. Subtract real taker fees on 1,600 trades and the green turns grey. On short-hold crypto scalping, the fee is not a rounding error — it's the whole result.
2. DCA dependence. The top of the leaderboard was stuffed with setups winning 95%+ of the time, but only because they averaged down on every loser. Strip the rescue buys and the equity curve inverts. A win that required tripling your risk at the bottom isn't a win.
3. Maximum drawdown. Plenty of "winners" sat through 14% adverse moves to bank a 1% gain. That math works in a spreadsheet with infinite margin and nowhere else.
Run those three filters and the 85 million collapses to a handful of candidates — and most of those still fail the moment you test them on data they've never seen. The honest count of clean, robust, fee-survivable crypto-scalping edges I found? Effectively zero.
So was it a waste?
No — because proving what doesn't work is worth real money. Every hour I don't spend live-trading a mirage is an hour my account survives. The brute-force search didn't hand me a money printer; it handed me a very long, very specific list of things to never risk capital on. That's a quieter result than "I found the holy grail," but it's the one that keeps you solvent.
The edges that have actually held up for me are boring and few: trend-following on the strong stuff, sizing by risk, and refusing leverage that turns a normal pullback into a liquidation. Before your next trade, size it honestly with the position size calculator and know your liquidation price first. The 85 million experiments mostly just confirmed that the simple, dull stuff is the only stuff that lasts.