217 of my 224 best-looking win rates all share the same hidden dependency

Our broad-screening stage logs two fields most people never see on a results screenshot: dca2_shadow_rate and dca3_shadow_rate -- how often a trade needed a second or third round of averaging down, modeled on the data but not always acted on live. I went looking for every config in our 2,331-row broad-screening batch with a win rate of 90% or higher, expecting a handful of standout strategies. I found 224 of them. Then I checked what they had in common.

96.9% of the near-perfect win rates share one rejection tag

Out of 2,331 configs that went through broad screening, 224 cleared a 90% win rate -- already a small, elite-looking group. But 217 of those 224, 96.9%, got tagged BROAD_REJECT_DCA3: the flag that fires when a config's dca3_shadow_rate runs too high. DCA3 is the deepest rescue level in our strategy families -- a third average-down add, used only after the first two already fired and the trade is still underwater. Across these 217 configs, the average dca3_shadow_rate was 48.6%. Not an edge case. Almost half of all logged trades needed the deepest possible rescue just to produce the win rate on the label.

44 survived it. 173 didn't.

Needing the deepest rescue half the time doesn't automatically mean a config loses money -- but on this batch it usually did. Of the 217 near-perfect win rates flagged for DCA3 dependency, only 44 (20.3%) still closed net positive. 173 (79.7%) had a profit factor under 1.0 and lost money overall, despite a win rate most traders would screenshot and post. The worst of the batch: a range_bottom_reclaim config on INITUSDT, 3,000 trades, a 90.6% win rate -- and a profit factor of 0.41, closing -$871.84 net. Three other configs on BLUAIUSDT, EDENUSDT and FFUSDT all cleared 90% win rate and still closed between -$659 and -$707. On the other side, the best survivor -- a range_bounce config on MEGAUSDT, also 3,000 trades, 96.9% win rate -- closed +$532.12 net with a 2.61 profit factor. Same rejection tag, same win-rate bracket, a $1,400 swing in outcome.

Same coin, same strategy, opposite result -- because of one parameter

PIEVERSEUSDT shows how thin the line is. One range_bounce config on it, 800 trades, 91.6% win rate, needed DCA3 on 75.1% of trades and closed at -$84.47. A different parameter set on the exact same coin and family, 3,000 trades, 96.6% win rate, needed DCA3 on only 61.0% of trades and closed at +$79.69. Same symbol. Same strategy family. A roughly 14-point gap in how often the deepest rescue fired was the difference between a loser and a winner. That's the same lesson I found looking at DCA1 last week, just inverted: heavier reliance on the first rescue level correlated with better outcomes across our full 12,331-config archive. DCA3 is a different animal -- it's the rescue you only reach after the first two already failed to save the trade cheaply, and leaning on it hard correlates with the opposite result.

It also means the win rate concentration isn't a one-coin story this time. 33 different symbols appear among the 217 flagged configs -- BILLUSDT (22 configs), ARCUSDT (17), BSBUSDT (16) and BEATUSDT (14) show up most, but no single coin dominates the way MEGAUSDT dominated our is_top_candidate flag. The pattern here isn't "one coin's price history got lucky." It's "a 90%+ win rate on this pipeline, almost regardless of coin, usually means the deepest rescue level is doing most of the work."

What I actually do with this

A win rate above 90% used to make me curious. Now it makes me check one field first: dca3_shadow_rate. If it's elevated, the win rate isn't describing the entry rule -- it's describing how often the system had to throw a third chunk of capital at a losing trade to flip it green before counting it. That's a survivability question, not a strategy-quality one, and it's exactly what the DCA survival calculator is built to answer: how far price can move against an averaging plan before the capital to rescue it runs out. Before trusting any win-rate headline -- ours or anyone else's -- I run profit factor next to it in the profit factor calculator. INITUSDT's 90.6% would have passed a win-rate-only filter. Its 0.41 profit factor is the number that actually tells you not to run it.

Method: 2,331 rows from our broad-screening stage (the same pipeline behind backtest-reports.html), covering five strategy families across live-tracked symbols with trades ≥ 20 for a fair per-config read. Win rate, profit factor, net, dca2_shadow_rate and dca3_shadow_rate are the pipeline's own logged fields, not re-derived after the fact.