"Flagship Expansion — Final Decision Report (2026-08-20)"

the decision metric (compare_mc8.md caveat).

TL;DR

Deploy the C59 blend (mc8 59-sym universe, h5+h21 model blend, region-A config). It keeps the full 20y mega-cap CAGR but fixes the recent regime that made the current B59 deployment painful:

The blend recovers essentially all of the h21-era edge (2026 +27% on the 51-sym A21, 2026 +3.9% at 59 syms) while keeping the mega-cap full-history return (B59 +1303%).

Overlay decision

Region-A stays. The held-out-validated overlay choice was already shipped to B59 and also validates on C59, but does not dominate region-A once blended:

• C59 region-A (ext=0.25 tl=0.00): full 14.34% CAGR, PF 1.44, 1y +10.2%, 2026 +3.9%. OOS test CAGR 17.71%. MC baseline +1310% (0/40 null beat it, p=0.000; P(loss)=0.000).
• C59 ext=0.25 tl=0.25: full 15.82% CAGR, PF 1.43, 1y +0.4%, 2026 -5.4%. OOS test CAGR 17.88% (best held-out), MC baseline +1716% (p=0.000; wider CI).
• C59 ext=0.60 tl=0.25: highest full CAGR (16.67%), OOS 17.80%, but 2026 +7.2% and OOS test DD 31.3% (vs 25.7%).
• C59 ext=0.40 tl=0.25: 16.22% CAGR, 2026 +10.5%, OOS 16.36% (mid-pack held-out).

Takeaway: the overlay edge that mattered on the h5 cache is compressed once h21 joins. Per the OOS discipline (compare_mc8 OOS chapter), do not hot-swap configs. Ship the blend at region-A; revisit the overlay only if a future OOS window supports it.

Validation summary (all on C59 blend)

What changed / artifacts

• scripts/build_mc8_h21.py — h21 cache builder (PREDICTION_HORIZON=21).
• scripts/watchdog_mc8_h21.py — auto-restart watchdog (build completed 59 syms, ~06:03).
• scripts/flagship_expand.py — sweep/MC/OOS tool (now handles --only59 correctly).
• Caches: phase1_..._w12_h21.pkl (new, 13.4 MB); blend built live in-memory.
• Reports: flagship_expand.md, flagship_expand_59.md, flagship_oos_split.md, flagship_oos_split_c59.md.

Deployment — DONE (2026-08-20)

The C59 blend is now the live flagship signal:

1. execution/flagship_signal.py — h21-tagged model checkpoints ({sym}_h21_{cutoff}.joblib, meta records horizon; h5 path unchanged for back-compat), parameterized extend_scores(horizon=), and blend_live() producing the traded cache phase1_..._w12_live_h5h21.pkl (combined = 0.5·h5 + 0.5·h21, ml_active = AND).
2. flagship_live.py update now runs h5 leg → h21 leg → blend; flagship_trader.py and all reporting tools read LIVE_NAME (the blend) unchanged.
3. h21 leg pre-warmed offline (56 periods, 224 new bars, 30 min) — daily updates are incremental (fetch once per leg, predict, no retrain).
4. Flagship loop restarted (new pid), start_bots.py --status → [ OK ].

Deployment requirements from the pre-approval analysis:

1. execution/flagship_signal.py runs daily extension for ONE horizon today (h5), and caches trained models at models/checkpoints/flagship/{sym}_{cutoff}.joblib. The h21 leg needs the same extension with PREDICTION_HORIZON=21 — model checkpoints must be horizon-tagged or the two legs overwrite each other.
2. Two live caches (..w12 and ..w12_h21) feed one blended combined = 0.5·h5.combined + 0.5·h21.combined with ml_active = h5 AND h21.
3. Recommend adding an h21 extension job to flagship_live.py daily flow (fetch once, extend both, blend) — single fetch, two model legs.

Files