Every One of My Top 50 Configs Blew Past Its Own Stop-Loss -- and Won

Two posts back I took the worst historical excursion (max_mae) on my top 50 candidates and turned it into a breakeven leverage number -- none of them clear 20x. This time I checked that same max_mae against a different column I'd been ignoring: final_invalidation_pct, the hard stop my engine is configured to cut the trade at. It should cap how far any of these 50 can ever drift against me. Every single row blows past it anyway.

The stops on these 50 configs sit between -3.50% and -4.00%. The actual worst price moves (max_mae) range from -5.00% to -14.24%. That's not a rounding gap. The ratio between the two -- how much wider the real move was than the stop that was supposed to cap it -- runs from 1.34x up to exactly 4.00x. All 50 out of 50 rows exceed their own stop. None of them exceed it by less than 1.3x.

By symbol, the gap isn't even close to uniform

MEGAUSDT -- the single cluster that scores highest in my whole top-50 list, 95%+ win rate, score up to 131.9 -- has the worst overshoot of the set, consistently 3.5x to 4x past its own stop, and every one of its 30 rows shows the exact same -14.24% worst excursion no matter what the stop parameter was set to. BILLUSDT, by contrast, barely clears its stop at all: 1.3x to 1.7x. Whatever is driving this, it isn't the stop setting. The stop varies row to row within a symbol. The actual worst move doesn't move with it.

The only way this reconciles

I went back through my own engine code on a related system before writing this, specifically how mae gets computed elsewhere in this pipeline: it's the lowest low after entry, measured against the original entry price, full stop. That's a fixed reference point. final_invalidation_pct is a position-level stop -- and these are DCA configs, every one of them with a dca1_pct leg that adds to the position at a lower price when the trade goes against me first. A stop measured against the post-DCA average cost basis is measured against a lower, moved reference point. The price can travel -14% from where I first clicked buy while sitting well inside -3.56% of the price I'm now actually carrying size at. Both numbers can be true at once. The stop didn't fail to fire. It was never looking at the same distance I assumed it was.

I checked whether this tracks cleanly with how often the DCA leg fires (dca1_rate) across all 50 rows, expecting a straight line -- more DCA, bigger gap. It doesn't correlate at all (r = -0.04 across the full set). BILLUSDT DCAs into 70%+ of its trades and still shows the smallest overshoot of the four symbols. MEGAUSDT and BILLUSDT both run the same family, range_bottom_reclaim, and land in completely different places. The gap is a property of how far each symbol's price actually travels after the DCA leg fills, not how often the leg fires. I don't have a clean one-line answer for why MEGAUSDT's floor is -14.24% and BILLUSDT's is -6.64% on near-identical config logic. I'm flagging it honestly rather than guessing a cause I can't check.

What this changes for sizing

If I'd sized leverage off final_invalidation_pct -- the number that's supposed to be the worst case -- I'd have been lowballing real risk by 1.34x on the gentlest row and 4x on the best-scoring one. The stop parameter describes what the backtest engine is configured to defend, not what the position actually lived through. For sizing, the number that matters is max_mae against the original entry, same as I used last time for the 20x leverage check, not the invalidation setting a DCA leg quietly moved the goalposts on. I run that worst-excursion number through the safe leverage calculator before touching the position size calculator, and I no longer treat a backtest's own stop-loss column as a ceiling on anything -- on this list, it wasn't a ceiling even once.