バックテスト

pv 0.3.1 = pv 0.3.0 + the portfolio cap inputs (engine build 0.2.1): 4 open positions / 1.0% open risk / 2.0% daily account loss / 3 per currency, evaluated by every instance on the account before an order. No trading value changed; identical to 0.3.0 on this window. Tier unchanged (proven). Headless run; generated ticks before 2026.

pv 0.3.0 = pv 0.1.0 + Japanese holiday calendar (engine build 0.2.0, InpSkipJpHolidays=true): no banks, no fix, so JP holidays are skipped and a gotobi flow that lands on one rolls back to the previous business day, the same way weekends already did. OUT-OF-SAMPLE 2010-2018; pv 0.1.0 on the same window: 469 / 1.26 / +$33 (trades / PF / net). About 30 holiday-day trades are replaced by ~30 rolled-back ones; the money is a wash (the holiday days had been carrying a few lucky outliers), the rule is now correct by construction. Headless run; generated ticks before 2026.

pv 0.3.0 = pv 0.1.0 + Japanese holiday calendar (engine build 0.2.0, InpSkipJpHolidays=true): no banks, no fix, so JP holidays are skipped and a gotobi flow that lands on one rolls back to the previous business day, the same way weekends already did. 2019-2026; pv 0.1.0 on the same window: 542 / 2.23 / +$132 (trades / PF / net). About 30 holiday-day trades are replaced by ~30 rolled-back ones; the money is a wash (the holiday days had been carrying a few lucky outliers), the rule is now correct by construction. Headless run; generated ticks before 2026.

Tier set: candidate (variant tested on one window; pv 0.1.0 is the live and proven one).

Tier set: PROVEN (PF >= 1.0 on 2010-2018 and 2019-2026, cause stated). Status: live.

Tokyo fix fade, first build. Hypothesis stated before the data was cut (fix-time flows, item 1 of the techniques-map article): USDJPY falls after the 09:55 JST fix on gotobi (5/10/15/20/25, rolled back) and month-end days. Short at the fix, exit on the clock 25 min later, 20-pip stop, no TP, one trade per day. The whole 2019-2026 window is the sample - no held-out window remains - so this is filed in-sample by construction; the evidence is the year-by-year consistency. Headless run; generated ticks before 2026 (7-8% real). | Live at 0.01 lot from 2026-09-10 22:12 JST as a forward test (candidate status kept). First admitted day: 2026-09-15.