Four Techniques Tested, One Survived: the Tokyo Fix

Yesterday's map of FX techniques ended with a list of four things we would test next, in order of how plausible the cause behind each was. This is the report. All four were run on the same data, OANDA's M5 and H1 bars for the majors, and every number below either sits on the presets page or in the working notes that accompany the public repository.

1. Fix-time flows: found, built, live

The Tokyo fix at 09:55 JST is the moment Japanese banks set the day's customer rate. The story everyone tells is that importers buy dollars into it, so USDJPY drifts up beforehand and gives it back after. On 1,991 trading days from 2019 to September 2026 the bars show only the second half of that story: there is no reliable drift up before the fix, but on the days when the flows are concentrated — the 5th, 10th, 15th, 20th and 25th (gotobi), rolled back to the prior business day when they fall on a weekend, and the last business day of the month — USDJPY and EURJPY fall after it.

The 20th, 25th and month-end carry most of it; the 5th and 10th add about a pip. On ordinary days the after-fix move is under half a pip: present, but not tradeable. The London 4pm fix, tested the same way on five pairs, showed only a month-end reversal on GBPUSD, about 90 events in seven years, which we recorded and did not build.

So we built an engine for the Tokyo one. TokyoFixEA sells at the first tick after 09:55 JST on an admitted day, closes on the clock 25 minutes later, carries a 20-pip protective stop and no take-profit, and trades once per day. It converts the broker's server clock to Japanese time itself, using the New York daylight-saving rule that the server follows, so a backtest and a live chart see the same moment. In the strategy tester, January 2019 to September 2026, 0.01 lot:

A variant restricted to the strong days (15th, 20th, 25th, month-end) raises the profit factor to 2.55 and the per-trade number by a sixth, at the cost of a third of the trades and about $30 of the total; the dropped days were weaker, not losing. Both versions are on the presets page. The full-day version is the one running live on USDJPY and EURJPY since 10 September, at minimum lot.

Two honest caveats. The whole 2019–2026 window is the sample, so there is no held-out period; the defence is that the hypothesis was on a published list before the data was cut, and that every full year agrees. And it is small money per trade, three to four pips on roughly seventy days a year per pair. What makes it different from everything else we have found is the cause: a participant who has to trade at a known time. That is why it held through 2019–2022, the years in which every one of our range-reversion presets lost.

2. Liquidity-sweep reversal: nothing there

The idea: price runs through a level where stops cluster, closes back inside, and reverses. We tested the Asian-session range in the London morning and the previous day's high and low through the day, on five pairs, 1,300 to 2,300 events per pair per level. At sixty minutes every mean sits between minus one and plus one pip, which is the spread, and the signs flip by pair and by year. Deeper sweeps did not help. Sweeps of yesterday's low on USDJPY leaned the wrong way. It also answered a question about our own system: the London-morning dip we trade is not a sweep effect. Buying a reclaim of the Asian low in that window is worth nothing; the money comes from the band distance, the RSI floor and the ADX ceiling.

3. Turn of month and day of week: one two-pip survivor, and a trap

This one taught us something before it gave an answer. The first pass, measuring each server day from open to close on bid bars, found the Asian session positive on every pair on every weekday with t-statistics up to 17. That is not a market effect. Spreads balloon at the daily rollover and over the weekend, which pushes the bid down at the day's open and at the Friday close, and that manufactures a positive Asia, a negative Friday and a positive Monday. Daily close-to-close returns carry the same artifact. We rebuilt the study on mid prices anchored at 07:00 server time, on H1 bars from 2010 for ten pairs.

Across sixteen years, split in two halves that had to agree, exactly one effect survives: the dollar rises on Tuesdays by about two basis points, which is one spread. The yen pairs' strong Monday, plus nine basis points since 2018, was minus four before 2018: a regime, not a calendar. Turn-of-month positions showed a few five-to-seven-basis-point cells with one half significant and the other not, on a dozen events a year, across eight positions tested.

4. Time-series momentum: reversion at every horizon

Lookbacks from one week to one year, holds from one to twenty days, ten pairs, 2010–2026, with and without skipping the last month. No cell is positive with a t-statistic of two in either half. Lookbacks under three months are negative at every horizon — the majors mean-revert at the weekly-to-monthly scale just as they do intraday — and the best cell in the grid is contrarian: fading the two-week move over the next twenty days, worth under a basis point a day. The one-year lookback is positive only in 2010–2017 and only on the yen crosses, which is the long yen trend, not a mechanism.

What the scoreboard says

With these four the continuation column of the map is closed on our own data: trend following, pullbacks, breakouts, daily-bias gating and multi-week momentum have all failed. Every positive result in this project is reversion at a known time and place, and the one result that holds across regimes has a participant behind it. That is the filter we will apply to whatever we test next: not "does this pattern exist" but "who has to trade here, and when".

Live today: four RangeRevert charts in the London morning (regime edges, minimum lot, with a kill switch) and two TokyoFix charts. All six are filed as candidates and their results will land on the presets page and the track record as they trade.

Nothing on this site is investment advice. Every result above is a backtest or a measurement on one broker's data, published so that it can be checked.