Ya he escrito sobre los dos extremos de este pipeline: la auditoría completa del archivo que encontró 9 pares en los que vale la pena confiar de los 1400 probados, y los 20 finalistas que todos fallaron en mi etapa más profunda de caminar hacia adelante. Esta es la parte intermedia: la bandera que uso para preseleccionar una configuración incluso antes de que llegue a esas etapas posteriores, y lo que sucede cuando realmente miras lo que preseleccionó.
La bandera se llama is_top_candidate. Es un pase de umbral en una carrera de puntuación anterior: ganancia neta, tasa de ganancias, factor de ganancia, nada más sofisticado que eso. De las 12.331 configuraciones puntuadas que abarcan 176 monedas diferentes, exactamente 50 lo han ganado. Fui a ver a qué monedas pertenecían esos 50, esperando un poco de propagación. No había ninguno.
50 banderas, 4 monedas
| Acuñar | Configuraciones marcadas | Tasa media de ganancias | AVG NET | Tasa media DCA1 |
|---|---|---|---|---|
| MEGAUSDT | 30 | 95.6% | $ 543.29 | 55,6 % |
| BILUSDT | 9 | 73.3% | 195,59 $ | 57.9 % |
| ARIAUSDT | 8 | 68.9% | $ 441.14 | 43,1 % |
| LIGHTUSDT | 3 | 35.1% | $ 331.57 | 63,5% |
| Otras 172 monedas probadas | 0 | — | — | — |
Cada fila de backtest_results con is_top_candidate=1, agrupada por monedas. 12.331 configuraciones totales puntuadas, 176 monedas distintas probadas.
Una moneda representaba el 60% de toda la bandera
MEGAUSDT solo representa 30 de las 50 banderas, cada una de ellas igual range_bottom_reclaim familia de estrategias con menor variación de parámetros. La mejor racha individual: 95.0% de tasa de ganancia, factor de ganancia 3.09, $ 668.08 netos sobre 1,600 operaciones, entrada dentro de un 1.10% de take-profit y un -0.81% de primer nivel de DCA. Eso no son 30 aristas independientes. Es el historial de precios específico de una moneda, anotó 30 formas ligeramente diferentes, y la bandera no pudo distinguir la diferencia entre eso y 30 ideas reales.
LIGHTUSDT es el extraño y vale la pena echar un segundo vistazo por sí solo: 35.1% de tasa de ganancias promedio, menos de la mitad de MEGAUSDT, y aún así despejó la bandera solo en la ganancia neta y el factor de ganancia (1.87). Una moneda puede perder más de lo que gana y aún así pasar un umbral construido en torno al factor neto y de ganancia. Esa es una forma real y defendible de una estrategia para ganar dinero, hasta que recuerdas que se suponía que el trabajo de la bandera era elegir las apuestas más seguras, no las de mayor varianza.
Volví a probar los 4 a través de la canalización más estricta. Los 4 fallaron
Since writing about the 20 finalists, I've been running coins through a newer broad-screening stage that scores on more than net and win rate — it also checks how P&L clusters across trading days. I went back and pulled these same 4 coins from that pipeline's results.
| Acuñar | Configs retested | Avg broad score | Best broad score |
|---|---|---|---|
| MEGAUSDT | 12 | -33.71 | 26.57 |
| BILUSDT | 24 | -99.72 | -87.46 |
| ARIAUSDT | 20 | -46.65 | -23.25 |
| LIGHTUSDT | 0 | did not survive to retest |
Same 4 coins, scored again through the newer 2,331-config broad-screening batch used in the 20-finalists piece.
BILLUSDT and ARIAUSDT scored negative on every single retested run — not just on average, their best run was still negative. MEGAUSDT, the coin with 95.6% average win rate under the old flag, barely cleared zero on its single best of 12 retested runs and averaged -33.71. LIGHTUSDT wasn't in the retest batch at all — it didn't survive whatever upstream filtering happens before a coin even gets scored again. None of the 4 made it into the 20 finalists I wrote about last week either. Two completely separate scoring passes over the same market data agreed on exactly nothing.
What I'm changing about how I read the flag
The flag isn't wrong on its own math — every one of those 50 configs really did clear its net/win-rate/profit-factor threshold on the data it was scored against. The problem is what the concentration hides. When 60% of a flag's entire output is one coin, it isn't telling you "this is a good strategy." It's telling you "this coin's specific 1,600-trade history happened to produce good numbers under these particular parameters." I can't tell those two things apart from inside the flag. I can only tell them apart by checking how many different coins produced a pass, and by that measure this flag was overfit before I ever got to walk-forward testing. BILLUSDT and ARIAUSDT both resurface again in a separate check I ran on every 90%+ win rate in the same broad-screening batch — BILLUSDT alone shows up in 22 of the 217 configs flagged for needing a third-level DCA rescue just to post that win rate.
Going forward I'm treating any flag or filter that clears fewer than a handful of distinct coins the same way I'd treat a single backtest with a suspiciously high profit factor — not disqualifying on its own, but a reason to check the concentration before trusting the number.
Before you trust a "top candidate" flag
- Count the distinct coins, not just the pass rate. 50 of 12,331 configs (0.4%) cleared my flag, but they came from only 4 of 176 coins tested. A low pass rate on a small handful of symbols is concentration risk, not selectivity.
- Watch for one symbol dominating the winners. 30 of 50 flags — 60% — were the same coin run with minor parameter changes. That's one price history scored many ways, not many independent edges.
- Re-score with a second, stricter pipeline before trusting a flag. All 4 coins that passed my original flag scored negative on average — or didn't survive to be retested at all — under a newer broad-screening pass.
- A high win rate on one symbol doesn't transfer. MEGAUSDT's 95.6% average win rate under the old flag became a -33.71 average score under the new one.
→ Calculadora de riesgo de ruina · → Calculadora de supervivencia DCA · → Calculadora de factor de beneficio · → Aprender
Preguntas frecuentes
What does a "top candidate" flag mean in backtesting?
In my pipeline it's a simple threshold flag on net profit, win rate and profit factor from an earlier scoring pass — not a validated result. Out of 12,331 backtest configs, only 50 (0.4%) ever earned it, and every single one came from just 4 of the 176 coins I'd tested.
Can a backtest filter overfit to a single coin?
Yes, and it's easy to miss if you only look at the flagged results and not at how concentrated they are. In this batch, 30 of my 50 "top candidate" configs — 60% of the flag's entire output — were the same coin, MEGAUSDT, run with slightly different parameters. That's not 30 independent ideas, it's one coin's price history scored 30 times.
Do backtest results survive being re-scored by a different pipeline?
Not reliably. I re-ran the 4 coins that had earned my old "top candidate" flag through a newer, stricter broad-screening pipeline. Two of the four averaged clearly negative scores, the coin with the highest original win rate barely cleared zero on its best run, and the fourth coin didn't survive to even be included in the retest.
Why did a 95% win rate coin still fail the retest?
Because win rate on one symbol's specific price path doesn't transfer. MEGAUSDT's 30 flagged configs averaged a 95.6% win rate and $543 net under the old scoring pass. Under the newer broad-screening pass, the same coin's average score was -33.71, and its single best run only just cleared positive.