spb/wp46_uqo

1# Feasibility report and Phase 2 pilot (2026-11-09)23## Data (after audit and panel construction)4- 42 instruments ingested from FinData OS (1-minute bars 2000/2008/2010/2013 → 2026-10-30; 230 M bars in the node cache),5  daily panels built with session windows D-02, realized measures D-03, validity rules D-05/D-06.6- Proxy sanity: mean(RV5ss + r_on²)/mean(r²) over valid days lies in [0.93, 1.17] for all ETFs, stocks, futures and FX7  (bond ETFs/ZN ≈ 1.16–1.24: the overnight share is small and intraday noise slightly inflates RV); crypto only from 20228  (BTC 1.24 on valid days — still noisier, disclosed).9- Validity rates: equities 0.75–1.00 (early-year thin trading in IEF, LQD, XLF, XLE, EEM, EFA), futures 0.89–1.00,10  FX ≈ 1.00, crypto 0.36 (BTC) / 0.48 (ETH) because of the 2022 cutoff.1112## Software verification13- 19 unit/integration tests, 5 Monte Carlo recovery tests (GJR, NAGARCH, Realized GARCH, HEAVY, MS-GARCH) and the14  rugarch 1.5-6 cross-check all pass; custom GARCH(1,1) reproduces arch's recursion to 1e-9 and arch's forecasts to 1e-15;15  all arch/custom log-likelihoods are within 0.3 of rugarch's maximum (FIGARCH: arch 1.47 below rugarch, documented).16- Leakage: a whole-job test perturbs post-origin returns and realized measures (×3 / ×9) and finds every forecast unchanged17  (GJR-t Tier A, HEAVY Tier B; and in the pilot GJR-t + HAR on real panels).1819## Pilot (validation period 2012-01-03 → 2015-12-31, window 1,000, refit 5 days; `results/metrics/pilot/`)20| instrument | origins | models | failures | fit time (s, mean) | notes |21|---|---|---|---|---|---|22| SPY | 1,006 | 9 | 0 forecast failures; GARCH-t 26/202 inadmissible refits (α+β ≥ 1) | 0.004–0.035 | QLIKE ratio to GARCH-N at h=1: GJR-t 0.88, APARCH-t 0.88, NAGARCH-t 0.89, EGARCH-t 0.93, HAR 0.99, Realized GARCH 0.85 |23| AAPL | 1,006 | 9 | 0 | 0.004–0.035 | GJR-t 0.92, EGARCH-t 0.94, APARCH-t 0.94, HAR 1.10, RealGARCH 1.00 |24| CL | 1,032 | 9 | 0 | 0.004–0.033 | HAR 0.73, GJR-t 0.94, EGARCH-t 0.96; at h=22 EGARCH-t 0.68 |25| EURUSD | 556 | 9 | EGARCH-t 20/556 forecast failures (simulation with inadmissible |β| ≥ 1 fits), GJR-t 23 inadmissible refits | 0.005–0.044 | HAR 0.89, GJR-t 0.92, NAGARCH-t 0.95; EGARCH-t 1.32 (failures excluded → operational penalty) |26Exploratory reading only (validation period): asymmetric models and realized-measure models beat GARCH(1,1) at h = 1 on27these four instruments; historical 22-day variance is 28–42 % worse; the levels-HAR direct projection is unstable for AAPL28(window includes high-variance episodes → large intercept), a known weakness that log-HAR addresses.2930## Compute budget (measured)31Per spec and instrument over 4 years: 0.5–10 s wall (MS-GARCH ≈ 0.08 s/refit, FIGARCH ≈ 0.26 s/refit, APARCH/EGARCH32simulation forecasts ≈ 2–3 ms/origin). Main run (3,276 jobs, OOS 2012–2026 ≈ 3,700 origins): ≈ 25–40 CPU-hours →33< 1 hour on the six nodes (116 workers). Robustness variants (expanding, window 2,000, refit 22) are affordable.3435## Risks and mitigations36- Inadmissible/near-integrated fits for t-distributed GARCH on FX and index ETFs: kept and flagged; failure-adjusted37  results reported; EGARCH simulation failures counted as failures (never replaced).38- Thin early-year intraday data for bond/sector ETFs: realized measures invalid on those days; common-sample analysis.39- Crypto: realized-measure evaluation only 2022–2026; Tier A daily returns from 2017 (BTC) / 2020 (ETH) with r² proxy as a diagnostic only.40- Vendor futures roll rule undocumented: ratio-adjusted continuous series guarantees no roll jumps in returns; RV on roll days spans two contracts (disclosed).4142Gate 2: **PASS**. Protocol frozen next (tag `prereg-frozen`), then the main run.43