I backtested my own trade-setup ranker across 5,000+ setups: no edge before fees, with any stop placement. What would you test next?
retained text
I run a few pay-per-call APIs for AI agents. The newest one, **trade setups**, scans 148 of the top-200 coins in one call. For each coin it computes six indicators (Ichimoku, RSI, MACD, EMA 50/200, Bollinger, OBV). Where they agree on a direction, it builds a plan: entry, stop (just beyond the nearest support or resistance), two targets and risk/reward. It returns them ranked. **Ranking:** signal strength × min(R/R, 3)/3, lower for coins more volatile than the median (× √(median ATR / ATR)), and lower when support or resistance sits between entry and target 1: the closer to the entry, the bigger the penalty. **Then I backtested it**, because I didn't want to sell this without knowing: **•** walk-forward, using the exact production code, only candles up to the ranking moment; **•** one ranking per day for \~4 months (121 days), 5,220 setups, each followed for 5 days (30 × 4h candles); **•** target 1 first = +R/R, stop first = −1R, stop and target in the same candle counted as a stop; no fees. **• 0.25 ATR beyond S/R (current):** 30% hit target · avg 0.00 R (top 5: +0.06 R) **• 1.0 ATR beyond:** 41% · avg −0.04 R (top 5: −0.03 R) **• 1.5 ATR beyond:** 45% · avg −0.02 R (top 5: −0.01 R) **• 2 ATR from entry:** 44% · avg +0.03 R (top 5: +0.18 R\*) **• 1.5 ATR beyond, target = 2R:** 23% · avg 0.00 R (top 5: −0.02 R) \*not robust: without trades above 5R it drops to −0.05 R. **Conclusion:** wider stops make you right more often, but you win less when you're right. Expectancy sits around zero for every variant. The indicators describe what *is* happening (trend, momentum, structure); over 5 days, in this period, they don't predict what happens next. So I'm positioning it as what it is: **a screener** that saves an agent 148 separate analyses and hands it the structure (levels, stop, targets). It's not a strategy. **What I'd like to hear:** **•** Which filter would you test first: only trading in the direction of BTC's trend, a longer horizon (10–20 days), or something else? **•** Is there a flaw in the method? For example, entering at the close of the signal candle, or counting a same-candle stop-and-target as a stop.
document fetched 2026-09-28 00:41 · canonical https://reddit.com/r/algotrading/comments/1wr9f3r/i_backtested_my_own_tradesetup_ranker_across_5000