Published
I keep a spreadsheet of the top candidates my GPU backtester spits out, ranked by score, and I usually scroll straight to win rate first — old habit, hard to kill. This week I noticed something that broke the habit for good. One coin, BILLUSDT, showed up nine times in the same batch. Same entry family (range bottom reclaim), same 808 trades every single time. And the win rate column read: 96.7%, 96.8%, 44.4%, 94.9%, 45.2%, 44.8%, 95.0%, 49.9%, 91.6%.
Nine configs. Same coin. Same trade count. Win rate bouncing between "almost never loses" and "loses more than it wins." If win rate measured anything real about a strategy, that swing shouldn't be possible on identical data.
What actually stayed still
Profit factor, on those same nine rows, moved from 1.68 to 2.12. That's it. A 53-point swing in win rate produced a 0.44 swing in profit factor. Net profit ranged +130 to +228 — all nine configs made money, all nine were within shouting distance of each other, and the ones with a coin-flip win rate weren't meaningfully worse than the ones that "almost never lost."
So whatever knob was turned between these nine runs, it wasn't buying or destroying edge. It was repainting the same rough outcome with a different win-rate label.
The knob was DCA-rescue, again
I checked the column that actually explains it: dca1_rate, how often a trade only closed green because it averaged down first. The two highest-win-rate configs (96.7%, 96.8%) had DCA-rescue rates of 72.3% and 71.7% — the highest I've logged on any coin, higher than the MEGAUSDT setup I wrote about a while back at 64.8%. The lowest-win-rate configs (44-50%) sat at 47.5-54.8% DCA-rescue. Still high in absolute terms, but a real, measurable gap.
Turn the DCA-aggressiveness parameter up, and a losing trade gets a second entry at a worse price, waits, and closes as a "win." Do that on 72% of your trades and your win rate climbs into the 90s. Do it less, and win rate falls into the 40s because more trades just... lose, honestly, on the first entry.
The part that actually surprised me
Here's the one I didn't expect. The single highest-net config in the batch — +228.2, profit factor 1.68 — was the 91.6% win-rate row with the lowest DCA-rescue rate of the high-win-rate cluster, 47.5%. The most DCA-dependent config, 72.3% rescue rate, actually finished with the lowest net of the bunch: +158.2. More doubling-down didn't buy more profit here. It bought a better-looking win-rate column and a worse bottom line.
Max adverse excursion stayed in a tight band too, -5.14% to -6.64% across all nine — tamer than MEGA's -14.2%, but not zero. At real leverage that's still a drawdown a fixed stop needs to survive, not a number a backtest can wave away.
What I do with this now
I stopped sorting my candidate list by win rate months ago. This is the reason I stopped looking at it as a tiebreaker too. On the same coin, same trade count, the same underlying edge produced win rates from 44% to 97% depending purely on how much the config was allowed to rescue itself with a second entry — and the version that rescued itself the most actually made the least money. If a strategy's headline number is win rate, ask what its DCA-rescue rate is before you ask anything else. Then check what the drawdown looks like at your leverage in the liquidation calculator, and model the averaging-down itself in the DCA reality calculator before you assume the "win" was ever really a win.