The Tokyo Fix: What It Is and How Our EAs Trade It

Every Tokyo business day at 09:55 Japan time, Japanese banks set the exchange rate they will use with their customers for the rest of the day. That moment is the Tokyo fix. On certain days of the month it leaves a measurable mark on the yen pairs, and it is the one trading mechanism, out of 28 we have studied, that we can stand behind.

This page is the reference for it: what the fix is, what we measured, the exact rule our EAs trade, and how those EAs are doing, live and in backtests. The numbers come from our own data and are published with their sources. How we found it is told in Four Techniques Tested, One Survived.

What the Tokyo Fix Is

At 09:55 JST the banks fix the day’s customer rate. Companies that need to pay or receive foreign currency often deal at that rate, so their orders arrive at a known time. The usual explanation is that Japanese importers have to buy dollars at the fix on settlement dates, whatever they think of the price.

Two kinds of day carry most of that flow:

• Gotobi days (五十日) — the 5th, 10th, 15th, 20th and 25th of the month, the traditional settlement dates in Japan. When one falls on a weekend or a holiday, it moves back to the previous business day.
• Month-end — the last business day of the month.

We cannot see those orders. What we can measure is what price does around the fix on those days, and that is what everything below is based on.

What We Measured

We used five-minute bars for USDJPY and EURJPY on 1,991 trading days, from January 2019 to September 2026. The popular story has two halves: the yen pairs drift up into the fix as importers buy, then give it back. Our data showed only the second half.

• Before the fix, no reliable drift. On gotobi days USDJPY moved −0.1 pips from 09:00 to 09:55, with a t-statistic of −0.1. That is nothing.
• After the fix, a fall. Selling at 09:55 on gotobi and month-end days and closing 25 minutes later averaged +3.7 pips on USDJPY (t 7.4) and +4.0 pips on EURJPY (t 8.2), over 549 days per pair.
• Every full year agreed. Seven of seven full years were positive on both pairs.
• Ordinary days are too small to trade. On other days the move after the fix was under half a pip, against a spread at the fix of 0.3 pips on USDJPY and 0.4 on EURJPY.

The days are not equal. The 20th, the 25th and month-end carry most of the effect; the 5th and the 10th add about a pip. Of the holding times we measured, from 5 minutes to two hours, 25 minutes gave the most reliable result on both pairs.

The move is small, and it is not smooth. Against the short, the price went a median 4.3 pips the wrong way before 10:30, and in the worst cases more than 50. A 20-pip stop was hit 40 times in 549 days on USDJPY without changing the average.

The Rule Our EAs Trade

The engine is TokyoFixEA, and its rule is short enough to read in full. These are the inputs from its source:

input int    InpFixHour          = 9;     // Tokyo fix 09:55 JST
input int    InpFixMinute        = 55;
input int    InpHoldMinutes      = 25;    // exit on the clock: fix + delay + hold (study peak: 25)
input TokyoFixDayFilter InpDayFilter = TFIX_GOTOBI_MONTHEND;
input int    InpSide             = -1;    // -1 = short (the measured effect), +1 = long (control)
input double InpStopPips         = 20.0;  // protective stop in pips (0 = none); MAE 90th pct was 16 pips
input bool   InpSkipJpHolidays   = true;  // no banks, no fix: skip JP holidays and roll gotobi flows back past them
input int    InpMaxTradesPerDay       = 1;

In words: on a gotobi or month-end day, sell at the first tick at or after 09:55 JST. Close on the clock 25 minutes later. Keep a 20-pip protective stop, no take-profit, and at most one trade a day. Skip Japanese holidays, because without the banks there is no fix.

The EA converts the broker’s server clock to Japan time itself. That conversion turned out to matter more than anything else on this page.

Backtests, and the Clock Error We Found

Our broker’s server clock has not always kept the same relationship to New York time. Before December 2014 it ran an hour differently, and until about 2020 it also slipped in the weeks when the US and Europe change their clocks on different dates. Our EAs assumed one fixed rule, so in parts of the 2010–2018 tests they traded an hour after the fix they were meant to trade. Live trading was never affected; the clock has matched since 2021.

We re-ran the older window on the true clock. The comparison is the honest test of the mechanism, because a real fix effect should get stronger when the test finally hits the fix:

USDJPY and EURJPY got stronger, which is what a real effect does. GBPJPY got weaker and still passes. AUDJPY’s older result almost disappeared — about one cent a trade — so we moved it from proven to candidate. It stays live on the strength of 2019–2026. The filed backtest reports on each EA’s page still show the old clock for 2010–2018; the true-clock figures above are from our re-run.

We also tested the same rule on other yen crosses. CHFJPY passes narrowly and stays a candidate. CADJPY failed its older window and was rejected. CNHJPY was rejected because the broker’s spread, about 1.2 pips, is larger than the fix move on that pair, about 0.6. The mechanism is there; after costs, it is not tradeable everywhere.

The Live Record So Far

As of 12 November 2026, the four fix charts have 17 closed live trades: 13 winners and 4 losers, for a net of +JPY 3,245. They now run at 0.03 lot. Every trade is on the live track record as it closes.

Seventeen trades prove nothing in either direction. The backtests above have several hundred trades per pair, and even they are a single broker’s history. Read the live line as a check that the EAs trade the rule as designed, not as evidence that the rule works.

What We Tried That Did Not Help

• Filters. We split the trades by everything we could see at 09:55: the move into the fix, the morning’s range, the overnight move, RSI, MACD, moving averages, the spread, the weekday. None separated the losing days from the winning ones in both halves of the data. Adding one would have improved the backtest and nothing else.
• The dollar after the fix, outside the yen. If importers buy dollars at the fix, the dollar might fall afterwards against other currencies too. We tested EURUSD, GBPUSD and the other dollar pairs in the Tokyo late morning. The recent half of the data showed nothing.
• Other fixes. The London 4pm fix, the ECB fix and the PBoC fix all have a clock, but their flows were mostly two-sided or already anticipated, and none survived. The London fix’s month-end fade on GBPUSD was built, then retired before its first trade when the clock correction pushed its older window to PF 0.72.

What This Does Not Tell You

• The cause is inferred. We see the price move, not the orders. The importer story fits the data; it is not something we can observe.
• It is small money per trade. Three to four pips, on roughly seventy days a year per pair. Spread and execution matter, which is why the backtests are labelled with their tick source.
• It can fade. “Proven” on this site describes the past. In the study, the partial 2026 year was slightly negative on both pairs over its first 49 days. If the flows change, the edge goes with them.
• One broker. Every figure here is on one broker’s data and clock. Yours may differ, especially in spread at 09:55.

What makes the Tokyo fix different from the 27 ideas that failed is the reason behind it: a clock, a participant who has to trade, and a known direction. That is now the first question we ask of any new idea. The rest of that story is in 28 Studies, One Survivor.

Nothing on this page is investment advice. Trading carries a real risk of loss.