Tokyo-Fix Gotobi Give-Back Short — 5m

Status: Regime entry: worked in one test window, not the other · Regime #6 · Listed 2026‑10‑14 Instruments: USD/JPY only · Timeframe: 5m · Direction: short only Frequency: ~59 trades/yr · Fees: 0.006%/side (about half a pip to a pip) · Needs the fill at the minute

Why it is in the regime section, not the library

It failed the six‑check bar. It is listed under the regime bar, which asks four things of one of the two test windows: a net profit after costs, still a profit without the best trade, a result above 95% of 400 random‑entry runs with the same exit, and at least 20 trades. It clears all four from January 2018 to October 2026 (507 trades), by a small margin: it made +3.3% where random Tokyo mornings with the same clock times had a median of −2.6%. From January 2010 to January 2018 it made more, but fails the third. Everything below the setup is disclosure, not a pass.

Where it fails the full bar:

• Random entry, 2010–17: +11.8% from January 2010 to January 2018, but random Tokyo mornings made nearly as much, a median of +8.5%, and it sits at the 79th percentile. The research log's reading: the give‑back after the fix is a feature of the Tokyo morning, not of gotobi days.
• Cross‑asset: EUR/JPY made +3.5% in 2012–17 but lost money without its best trade, and lost 0.3% in 2018–26. AUD/JPY made +10.3% in 2010–17, no more than random mornings did (77.3rd percentile), and lost 1.6% in 2018–26.

It is the other half of the morning entry #7 trades. Japanese importers settle on gotobi days and buy their dollars at the 09:55 Tokyo fix; entry #7 is long into the fix, and this is short for the two hours after it. Both came out of sweep #13, pre‑registered on 2026‑10‑08, which tested the two legs on USD/JPY, EUR/JPY and AUD/JPY in two windows. That is 12 window tests. With a 95th‑percentile bar, about one window test in twenty clears by luck alone. Seven cleared: all six of the long into the fix, and this one.

The setup (compilable spec)

Market: forex USD/JPY, 5m trading timeframe. Shorts only. Day: a gotobi day (gotobi_day == 1): the 5th, 10th, 15th, 20th, 25th or 30th, moved back to the previous Japanese business day over weekends and holidays. Tuesday to Friday only (day_of_week 2–5). Entry: short on the 5‑minute bar that opens 00:50 UTC and closes at 09:55 Tokyo, the fix (hour_utc ≥ 0.83 and hour_utc < 0.9). Exit: the bar that opens 02:55 UTC and closes at 12:00 Tokyo (hour_utc ≥ 2.91 and hour_utc < 3). No stop, no target. Costs: 0.006%/side · no rollover (it holds through none).

Tokyo keeps no daylight saving, so the UTC times never move. At 00:50 UTC the UTC date is the Tokyo date, so day_of_week (ISO, 1 = Monday) reads the Tokyo weekday. Run through the app's engine on the same candles, this config reproduces the research harness trade for trade in both windows.

Evidence: two windows on USD/JPY five‑minute bid candles

Returns are per‑trade percentages summed, unlevered, like every return in the research log. In 2018–26 that came to 0.84 pips a trade after costs; in 2010–17, 2.3. Holding USD/JPY made +21.1% and +39.9% over the same windows; this trade is short.

• Random mornings, 2018–26: 400 runs of the same clock times on random eligible days (Tokyo Tuesday–Friday business days, gotobi or not) had a median of −2.6% and a 95th percentile of +1.6%.
• Random mornings, 2010–17: a median of +8.5% and a 95th percentile of +14.9%, against the real +11.8%.
• Data: Dukascopy's five‑minute bid candles to December 2011, FXCM's archive from January
  2012. FXCM has no quotes from late April to late July 2026 and for a few weeks in 2024–25: 4 gotobi days in 2010–17 and 29 in 2018–26 have no 09:55 bar, and they drop out of the random mornings too.

The record by year

*2026 through September, with no data from late April to late July (16 gotobi days missing). Each trade counts in the year it was entered.

• Small and spread out: in 2018–26 no quarter made more than +1.4% and no trade more than +0.51%. Split where the gotobi paper's sample ends, 2018–20 made +1.6% and 2021–26 +1.7%. In 2010–17 one trade, on 24 June 2016, made +3.86% of the window's +11.8%.
• Since 2018 it lost money in 2021, 2023 and 2026 so far, each time by 1.2% or less.
• Last high: May 2025, at +16.7% summed since January 2010. It ended September 2026 at +15.1%, 1.6 points below.
• Last twelve months (October 2025 to September 2026): 45 trades, −0.7%.

What the result depends on

No cost‑stress, late‑fill or neighbor run was made for it: sweep #13 registered those for a strategy that cleared the full bar and ran them only for the long into the fix (research log, sweep #13, G1 “Robustness” and “Neighbors”).

Before you adopt it

• It is small — +3.3% over 507 trades in 2018–26, under a pip a trade — and it assumes fills on the minute at the 09:55 fix and at noon, at 0.006% a side.
• In 2010–17 it made more, +11.8%, but random Tokyo mornings made nearly as much (79th percentile).
• No stop. Its worst trade lost 1.79%, on 15 September 2010; a surprise in the Tokyo morning hits it in full.
• Paper‑trade it first, with an order working at the fix rather than an alert acted on later. Adopting it watches USD/JPY on the 5‑minute chart and alerts every setup, about 59 a year.

Every number here can be re‑run: npx tsx scripts/research/sweep13-tokyo-fix.ts for the windows and the null. The record by year, the last twelve months and the last high are the same runs' trades, grouped by the date they were entered.