β fresh β 0.0h ago Β· Counterfactual replay of 26543 closed paper trades under each exit rule β reuses the paper ledger, no new data. Peak-based estimate (armedβtrail from peak; unarmedβnatural exit; floored at hard-SL; peak clamped 50Γ). Directional, not a tick-exact backtest. Generated 2026-11-19 20:24 UTC.
β οΈ 260723 finding: this replay does NOT reliably predict live results β TRAIL-hair and TIGHT5 both looked strong here and both lost ~24-25% when actually run live (see paper.html LIVE STRATEGY badge for the real number). Treat rankings on this page as a shortlist to test, never as promotion proof by itself.
π― Recommended A/B mix
40% SCALP_TRAIL + 33% TIGHT5 + 27% SCALP
Weighted by profit Γ consistency β (expectancy Γ win-rate)^1.5 over the top-3 positive-expectancy exits, so the best exit (SCALP_TRAIL) takes the lionβs share and losing exits get zero. Pure SCALP_TRAIL maximizes raw profit (Row SCALP_TRAIL below: $914,828); this blend concedes a little for robustness so no single rule owns 100% (regime / overfit hedge). Row MIX shows the blended result vs pure ACTUAL.
Optimized on first 70% of days, champion must beat the incumbent on the LAST 30% out-of-sample (OOS) β this is still the SAME peak-replay math as above, so OOS alone is not live proof. Recs only β any live change needs Chris's approval AND a genuinely independent forward test (design pending, see project_exit_optimizer). S=single trail Β· T2=TP+trail Β· T3=TP+trail+moon-bag.
Legend:β gold number = stable β₯6 consecutive runs (same champion each re-run). π¨ orange-outlined row = this segment hit SIREN tier (stable ~7d AND its shadow-lane number beat actual) β historically this has STILL not meant "safe to promote" (see the warning above), it just means the candidate is worth a real forward trial.
SEGMENT
CHAMPION EXIT
TRAIN
OOS
OOS 95% CI
INCUMBENT OOS
N
STABLE RUNS
GLOBAL
S(arm8/tr15/sl-20) π’
+30%
+18%
[+12%,+22%]
+5%
9384+1536
16β
regime:poor
S(arm5/tr15/sl-20) π’
+39%
+24%
[+16%,+35%]
+9%
5000+809
40β
regime:ground
S(arm10/tr15/sl-20) π’
+22%
+12%
[+1%,+16%]
+0%
3676+426
11β
profile:SCALP-SWEEP-REV
S(arm30/tr15/sl-20) π’
-1%
+3%
[-3%,+13%]
+1%
2848+534
194β
chain:robinhood
S(arm5/tr15/sl-20) π’
+67%
+8%
[+3%,+13%]
-6%
2605+256
55β
profile:SWING_ARM21
S(arm5/tr15/sl-20) π’
+58%
+24%
[+16%,+37%]
+6%
2346+442
264β
chain:bsc
S(arm10/tr15/sl-20) π’
+19%
+23%
[+13%,+32%]
+6%
2160+326
62β
src:volume_path
S(arm15/tr15/sl-20) π’
+3%
+8%
[+3%,+14%]
-2%
2127+319
5
chain:sol
S(arm15/tr15/sl-20) π’
+26%
+21%
[+6%,+32%]
+9%
1888+463
2
profile:SWING_SL25
S(arm8/tr15/sl-20) π’
+44%
+16%
[+6%,+28%]
+3%
1155+528
13β
chain:eth
S(arm10/tr15/sl-20) π’
+13%
+60%
[+32%,+88%]
+36%
1376+168
39β
src:whalemover_rh
S(arm5/tr15/sl-20) π’
+36%
+9%
[+3%,+17%]
-7%
1009+107
25β
regime:extreme_poor
S(arm5/tr15/sl-20) π’
+13%
+8%
[+3%,+16%]
-3%
864+145
11β
src:source_strategy
S(arm15/tr15/sl-20) π’
+18%
+55%
[+4%,+98%]
+39%
711+117
125β
src:wallet_intel
S(arm5/tr15/sl-20) π’
+30%
+20%
[+9%,+25%]
-1%
486+123
195β
src:V
S(arm5/tr15/sl-20) π’
+110%
+47%
[-15%,+114%]
+15%
500+15
1
src:sa_accum
S(arm5/tr15/sl-20) π’
+19%
+20%
[-1%,+34%]
+3%
427+63
264β
src:whalemover
S(arm10/tr15/sl-20) π’
+4%
+10%
[-4%,+20%]
+1%
380+54
77β
src:ranker
S(arm5/tr15/sl-20) π’
+58%
+49%
[+23%,+78%]
+22%
272+53
61β
src:clu:fomo
S(arm5/tr15/sl-20) π’
+10%
+6%
[-1%,+17%]
-2%
252+65
2
src:alpha_brief
S(arm5/tr15/sl-20) π’
+20%
+35%
[+11%,+68%]
+16%
198+100
13β
profile:TIGHT5
S(arm15/tr15/sl-20) π’
+48%
+70%
[+45%,+89%]
+43%
195+22
6β
src:A
S(arm15/tr15/sl-20) π’
+31%
+15%
[-7%,+32%]
-3%
159+28
22β
src:D
S(arm15/tr15/sl-20) π’
+17%
+18%
[-6%,+53%]
+1%
116+23
2
src:sniffer
S(arm30/tr15/sl-20) π’
+6%
+7%
[-0%,+13%]
-2%
68+25
2
src:whale_driver
S(arm10/tr15/sl-20) π’
+0%
+22%
[-2%,+53%]
+7%
61+15
16β
src:wallet_goldilocks
S(arm10/tr15/sl-20) π’
+41%
+57%
[+7%,+114%]
+32%
48+19
3
src:oggy_kol
S(arm8/tr15/sl-20) π’
+186%
+52%
[-2%,+71%]
+25%
40+15
2
Engine next steps:
β οΈ UNCONFIRMED champions look strong in eventual-peak replay but have NOT cleared a path-true randomized actual-fill test (TRAIL-hair and TIGHT5 were previously promoted from replay and both lost live, see since_promotion on paper.html): profile:SWING_ARM21:S(arm5/tr15/sl-20)(tier=NO_PATH_TRUE_AUTHORITY), profile:TIGHT5:S(arm15/tr15/sl-20)(tier=NO_PATH_TRUE_AUTHORITY), chain:sol:S(arm15/tr15/sl-20)(tier=NO_PATH_TRUE_AUTHORITY), chain:bsc:S(arm10/tr15/sl-20)(tier=NO_PATH_TRUE_AUTHORITY) β replay may nominate only; do not promote
grid-edge params in GLOBAL, profile:SWING_ARM21, profile:SWING_SL25, profile:SCALP-SWEEP-REV β widen ARM/TRAIL grid next run
thin OOS windows (profile:TIGHT5, src:oggy_kol, src:A, src:D) β keep collecting, re-validate in 3d
ENTRY quality (median peak): best M1 +53% (n273) vs worst known_conv -1% (n1256) β exits can't fix bad entries; consider demoting the worst source in EDGE LAB
tranche verdict is per-segment (see champions) β wire winners as SHADOW lanes in paper_universal for live A/B once approved (EP queue)
Exit strategy leaderboard β uniformly applied to every trade
Click a row to load its daily / weekly / monthly heatmap below. Expectancy = mean per-trade return; the tail (avg-win, p-return-x) matters more than win-rate for these power-law payoffs.