# Gotobi round trip

The edge lab's full test of one rule, as it was scored. [How we test](/concepts/research/.md) explains each part.

Edge-lab test · hypothesis 64 · written 2026-11-04 · scored 2026-11-04 Pass

## The rule

The rule:

* Gotobi days: the 5th, 10th, 15th, 20th, 25th and last day of each month; when Tokyo is shut, the Tokyo business day before, from the Tokyo Stock Exchange's calendar.
* Tuesday to Friday only: Monday's 03:00 Tokyo is before the week's open.
* Long USDJPY from the first minute at or after 03:00 Tokyo to the first at or after 09:55, then short from there to the first minute at or after 12:00 Tokyo.
* One value a day: the two legs' log returns after costs, summed, over USDJPY's daily volatility.

### Pass mark, written before the test

Pass: mean above zero, t > 3, and beating random in at least 99% of 10000 runs (seed 1). Random keeps the round trip (long 6 hours 55 minutes, then short 2 hours 5 minutes) and starts it at a random five-minute mark of the same month whose fills have quotes, each fill paying the profile's spread at its own minute.

### Verdict

Passes all four of goal.md's tests, and is working now. The round trip makes +0.095 of a day's volatility a day after the broker's own costs on the years where its short leg was unread (t 4.77), beats every one of 10000 random round trips, holds on a second feed (HistData t 4.63), and makes nothing on other days. Its fresh holdout, four yen crosses never used on gotobi, passes once at t 3.37 in both halves, three of the four pairs alone above t 3.5. Ten of its eleven years are above zero, the latest included. One trade a day, so the day's score is the trade's. It goes to the demo (goal.md) beside hypothesis 62's leg, earns no live money or product claim until the demo matches the scorecard, and the months after 2026-09 are its forward record. The cause is the same as hypothesis 62's, so they are one lead, not two.

Values are in a day's volatility after the broker's costs, one value a day, unless a column says pips.

Traded by [AIMS Gotobi EA](/eas/aims-gotobi/.md). Its [Evidence](/eas/aims-gotobi/.md#evidence) section shows this test beside the tester runs.

## Against random

Random entries keep the same side, stop and target. The holdout was kept back and scored once, on the fixed rule.

| Span                                                                       | Days | Value a day | t    | Random mean       | Beats random      |
| -------------------------------------------------------------------------- | ---- | ----------- | ---- | ----------------- | ----------------- |
| USDJPY, OANDA, 2016-01 to 2026-09                                          | 577  | +0.095      | 4.77 | -0.0371           | 100% of runs      |
| USDJPY, HistData, 2016 to 2025-03                                          | 463  | +0.1039     | 4.63 | No random control | No random control |
| EURJPY GBPJPY CHFJPY NZDJPY, OANDA, 2004-05 to 2026-09Holdout, scored once | 1259 | +0.0478     | 3.37 | -0.0806           | 100% of runs      |

## Robustness checks

Each check as it was written before the test.

| Check                                                                        | Result                                  | Holds |
| ---------------------------------------------------------------------------- | --------------------------------------- | ----- |
| both halves are above zero                                                   | First half +0.1271, second half +0.0631 | Yes   |
| the short leg alone is above zero                                            | +0.0286 a day, t 2.61, 577 days         | Yes   |
| twice the costs is above zero                                                | +0.0572 a day, t 2.88, 577 days         | Yes   |
| the same round trip on Tuesday to Friday days that are not gotobi makes less | -0.0265 a day, t -1.69, 1271 days       | Yes   |
| HistData's months are above zero                                             | +0.1039 a day, t 4.63, 463 days         | Yes   |

## Each market

Every market scored, weak ones included.

| Market | Value a day | t    | Days |
| ------ | ----------- | ---- | ---- |
| USDJPY | +0.095      | 4.77 | 577  |
| CHFJPY | +0.064      | 4.35 | 1180 |
| EURJPY | +0.0537     | 3.52 | 1189 |
| GBPJPY | +0.0612     | 4.13 | 1220 |
| NZDJPY | +0.0123     | 0.63 | 1164 |

## By year

The main market, one value a year.

* 2016+0.0535
* 2017+0.1658
* 2018+0.2123
* 2019+0.0609
* 2020+0.1002
* 2021+0.1171
* 2022+0.1548
* 2023-0.011
* 2024+0.0695
* 2025+0.0572
* 2026+0.0224

## Costs

* Spread

  PepperstoneUK-Demo, the median spread at each fill's server minute

* Commission

  0.007% a leg

* Financing

  1% a year markup for the night held

* Round trip paid

  2.66 pips

* Largest round-trip cost that keeps t above the bar

  5 pips

* Largest round-trip cost that keeps the mean above zero

  9.2 pips

## Tester tie

Each tester run of AIMS Gotobi EA compared with the lab's scorer, day by day. It shows the EA trades the rule; it is never the verdict.

| Run                               | Scorer days | Days the EA traded | Extra days | Latest open (minutes) | Days compared | Correlation | Mean difference (pips) | Matches the scorer |
| --------------------------------- | ----------- | ------------------ | ---------- | --------------------- | ------------- | ----------- | ---------------------- | ------------------ |
| real ticks                        | 48          | 97.9%      | 0          | 0                     | 33            | 1           | 1.72                   | Yes                |
| every tick generated from M1 bars | 668         | 99%        | 7          | 0                     | 47            | 1           | 1.95                   | No                 |

every tick generated from M1 bars: The EA applies Tokyo's holiday law as it stands since 2022 to every year; the scorer reads each year's exchange calendar. Where a gotobi day's shift crossed a holiday that has since moved, they part. No day after 2021 differs, so the EA's days from 2022 are the scorer's.

[Raw lab export (JSON)](/reports/aims-gotobi/lab/h64.json)