← all research
REJECTED

"Multi-Seed Holdout Synthesis — 60 random S&P names, 3 independent seeds"

2026-08-12

Pipeline identical to the flagship (enhanced features, regression target, top-28, 12mo retrain, 50-bar embargo, region-A config, 20y daily). All names excluded from the flagship stride-10 universe. Live breaker = 0.50; 0.35 shown for contrast.

Table

SeedNames (ex)Total (0.50)CAGRMaxDDPFTradesWR2023-26 (0.50)2023-26 (0.35)
42ARES, ATO, AVY, BLK, CMG, GNRC, GOOGL, HLT, IVZ, JBL, MRNA, NVDA, PAYX, PGR, PNW, PRU, RMD, STX, SWK, WDC+396.9%+8.46%39.3%1.2411,52746.2%+91.5%+0.0% (dead)
7ABNB, AME, BG, BNY, CPRT, DOC, DUK, HSY, INVH, LII, MDLZ, NFLX, PRU, RTX, SMCI, SNDK, TEL, TJX, TXT, UBER+583.9%+10.22%43.7%1.2581,41048.4%+38.3%+0.0% (dead)
1337(see report)+283.4%+7.04%40.1%1.2481,41746.2%+73.1%+73.1% (never locked)

SPY 20y: +757.9% | SPY last 5y: +85.3%

Verdict

  1. The cross-stock edge is real and stable: 3/3 independent 20-name samples profitable with PF 1.24-1.26 (avg 1.249, spread 0.017) and total +283%..+584% over 20y. This is not sample luck -- the strategy transfers across stocks.
  2. The edge is modest vs buy&hold: every seed underperforms SPY's +757.9% absolute. On unseen names the strategy behaves as a low-beta long-only timing engine, not an alpha monster (+283%..+584% in a market that 5x'd).
  3. The 0.35 drawdown lock murdered the last 5 years on 2 of 3 seeds: seed 42 went +0.0% / 0 trades in 2023-26 (vs +91.5% lifted); seed 7 +0.0% (vs +38.3%). The 0.50 breaker fixed the deadlock; seed 1337 never locked.
  4. Universe effect persists: flagship live bundle (51 names) PF 1.46 vs ~1.25 on unseen names -- stronger names, stronger results. The model is stock-agnostic; the names are not.

Bottom line

The flagship edge generalizes to never-seen stocks with stable P/L ~1.25 (3/3 seeds), beats the market on the flagship's own universe (PF 1.46), and needs the 0.50 breaker to stay alive through drawdowns. Remaining question: can the pooled zero-shot model (training in progress, 13/19 cutoffs done) lift the unseen-name PF above the per-symbol 1.24-1.26 baseline.