Published
Out of 12,331 rows in my backtest table, 50 clear the bar my pipeline uses to flag a config as worth testing deeper -- score, win rate and sample size all above threshold. I pulled all 50 and looked at a column I don't usually lead with: max_mae, the worst percentage the position moved against entry before the trade closed. Then I converted each one into a breakeven leverage number -- the highest leverage where that drawdown stays short of the liquidation line, ignoring fees and funding, which only makes the real number worse. Not one of the 50 clears 20x. 38 of them don't clear 10x.
The breakeven math is blunt on purpose: liquidation at leverage L happens roughly when price moves 100/L% against you, before maintenance margin tiers or funding make it worse. A config with a worst drawdown of -10% has already eaten the full move 10x leverage allows. Run it at 10x and the position is gone before the backtest's own win shows up. I'm not applying a safety margin here, just the raw arithmetic -- and the raw arithmetic alone is enough to fail every single one of these 50 rows above 20x.
Four symbols, one surprise
The 50 top candidates aren't 50 different ideas. They're four symbols, all the same range_bottom_reclaim family, split across parameter variants:
| Symbol | Rows | Avg win rate | Avg score | Worst max_mae | Breakeven leverage |
|---|---|---|---|---|---|
| MEGAUSDT | 30 | 95.6% | 121.8 | -14.24% | ~7.0x |
| LIGHTUSDT | 3 | 35.1% | 49.1 | -10.18% | ~9.8x |
| ARIAUSDT | 8 | 68.9% | 48.4 | -13.28% to -5.00% | 7.5x-20.0x |
| BILLUSDT | 9 | 73.3% | 49.4 | -6.64% to -5.14% | 15.0x-19.4x |
MEGAUSDT is 60% of the entire top-50 list. It also has the highest average win rate of the four (95.6%), the highest average score (121.8, against 48-49 for the other three) and the highest average net result per config (543 units versus 196-451 for the rest). By every metric my pipeline sorts on, MEGAUSDT is the best thing in the table. It's also the cluster with the worst breakeven leverage of the four -- 7x. The config my own scoring system likes most is the one that forgives leverage the least.
Same symbol, different pain
ARIAUSDT shows the other half of this. Its 8 rows span a worst max_mae of -13.28% down to -5.00% -- a 3x spread in how much drawdown you'd sit through, inside the same symbol and same strategy family, just different parameter choices (DCA rescue threshold, time-stop window). The score doesn't know the difference. Two ARIAUSDT rows can post nearly identical score and win rate while one of them would liquidate at 7.5x and the other survives to 20x. Sorting by score alone picks the profit. It says nothing about which specific row you're getting.
What I actually do with this
I stopped treating "top candidate" as a leverage recommendation. It isn't one -- it's a profit/sample-size filter that has nothing to say about drawdown tolerance. Before I'd run any of these 50 live, I pull its max_mae, convert it to a breakeven leverage the way I did above, and then run that number through the safe leverage calculator with a real buffer under it -- not the breakeven line itself, which is the exact leverage where I'd get liquidated on the worst historical excursion with zero margin for a worse one next time. I also check the resulting liquidation price against the mark price liquidation gap calculator, since mark price and last price can diverge enough on thin pairs like these four to eat the buffer I thought I had. Score tells me which configs made money in the sample. It took a second, separate pass through max_mae to find out which of them I could actually survive holding, the way I wrote about when the single highest score in the whole 12,331-row table turned out to be a formula artifact, not a real result. A row clearing the top-50 filter answers one question. Surviving the leverage you'd want to run it at is a different question, and nothing in the score answers it for you.