Out-of-sample track record

period2025-01 → 2026-05 selectionexpanding top-1% winTP-first beats SL-first live TP / SLL 0.80 / 0.60 · S 0.50 / 0.40
LONG
58.8% win rate
857W600L
P-35 (LSTM lb96)
TP +0.8% / SL −0.6% · edge +8.8pt
SHORT
59.2% win rate
251W173L
P-14bv3 (LSTM lb48, bull-filter N7/X+5%)
TP +0.5% / SL −0.4% · edge +9.2pt
COMBINED
58.9% win rate
1108W773L
1,881 resolved signals
edge +8.9pt vs random

Skill & risk — beyond the headline win rate

1,881
sample size (and growing)
+16.1pt / +14.3pt
selection edge L / S
108 / 48
max drawdown L / S (calls)
108 / 30
longest losing streak L / S (calls)
Selection edge Win rate of the top-1% signals minus the win rate of taking every bar (baseline L 42.7% · S 44.9%). It isolates how much the model's ranking adds.
Drawdown & streak Max drawdown = the deepest peak-to-trough of the Σ (win − loss) curve; longest losing streak = the most consecutive SL-first calls. Both counted in calls, never in money.

Cumulative correct calls (Σ win − loss · unitless)

Calibration — does higher conviction mean higher accuracy?

Signals grouped into expanding-window conviction tiers — the same live-honest ranking the model applies. Lower tiers sit near the coin-flip baseline; the edge is concentrated in the extreme tail. The outlined top 1% bar is exactly the slice the model selects, and it clears 50% on both sides.

LONG · win rate by conviction band
SHORT · win rate by conviction band

Monthly win rate (top-1%, decided)