The Anomaly Killed by Its Own Paper

This Anomaly Has a Date of Death

Trading edges usually fade quietly. They get worse in some years, then in others, and at some point it turns out they have been gone for a while. You can argue about the reasons forever, because nobody saw the moment it broke.

The Gotobi anomaly in USD/JPY is a rare exception. It has a date, and that date coincides with the publication of the academic paper that described it.

Where the move came from

Japanese importers pay foreign suppliers in dollars. The rate for those payments is not taken from the market at the moment of the transaction - it comes from the daily TTM benchmark that banks fix at 9:55 Tokyo time. On top of that sits a business custom called Goto-barai: settlements are timed to days divisible by five - the 5th, 10th, 15th, 20th, 25th and 30th.

The mechanics follow from there. A bank knows about its clients' dollar demand in advance, before 9:55, and covers it gradually in the preceding hours rather than at the fixing itself. The demand is asymmetric: Japanese exporters, who receive dollars, have no Goto-barai custom, so no offsetting sell flow appears on the same days. The result is predictable upward pressure on USD/JPY into 9:55 on days divisible by five, and a pullback once the benchmark is set and the demand is met.

This matters because it is not a chart pattern. It is a flow of real payments on a known schedule.

This is not Japanese exotica

The same mechanism has been documented far more broadly. Krohn, Mueller and Whelan, writing in the Journal of Finance (2024), study the nine most traded currency pairs from 1999 to 2019 and show that the dollar systematically appreciates ahead of three currency fixes - Tokyo at 9:55, Frankfurt (the ECB fix) at 14:15 and London at 16:00 - and depreciates after each one. Over 24 hours this traces a repeating pattern of reversals.

The cause, according to the authors, is not information but inventory. Clients present an unconditional demand for dollars at the fixes, dealers absorb that risk and raise their quotes, then unwind afterwards. The average size of the effect is about two basis points on the dollar portfolio.

What makes the paper convincing is less the numbers than three natural experiments. Japan does not observe daylight saving time, so measured in New York hours the dollar's peak around the Tokyo fix shifts by exactly one hour when US clocks change, following Tokyo rather than the calendar. On the 304 Japanese holidays in the sample, when the benchmark is not published, the pattern around Tokyo time disappears while the European reversals stay intact. And before the London (1994) and ECB (1999) fixes were introduced, these turning points did not exist in the data at all. The reversal is tied to the existence of a published reference rate, not to the hour of the day.

What the paper claimed

In January 2023, Bessho, Sugimoto and Suzuki published work specifically on Gotobi. Using 2018-2020 data (185 days divisible by five), they tested two hypotheses. First: buy the dollar at 3:00 Tokyo time and sell at 9:55. Second: sell right after 9:55 and close by noon, capturing the reversal.

Their results, already net of the spread: the first hypothesis produced a profit factor of 1.46 with a 60% win rate; adding a moving-average crossover filter took it to 2.62 and 68%; the second hypothesis gave 2.09; both together, 2.60.

The most valuable part of the paper is none of that. The authors ran the same mechanics on days not divisible by five and got a profit factor of 0.51 with a 41% win rate. Without the calendar, the identical setup loses convincingly. This is precisely the test most published strategies fail: their rules make money wherever you point them, which means the money is coming from curve fitting rather than from the described mechanism.

What the data showed after publication

We tested both hypotheses on minute-level USD/JPY history from 2020 to 2026. The figures below are gross, before costs: the goal was not to evaluate a trade but to see whether the move exists at all.

The tables carry a "t-statistic" column, and without it the other numbers mean little. It shows whether a result is distinguishable from luck: the average result divided by its own margin of error, in other words how many measurement errors it sits away from zero. The rough rule: above two in absolute terms, luck is a hard explanation; below two, an easy one. A profit factor of 1.62 sounds equally solid on two hundred trades and on five. Only this column tells them apart.

The 2020-2022 window is adjacent to the authors' sample and serves as a check that we understood the rules correctly:

That is close to what the authors published on their own sample (1.46 and 2.09). So we read the rules correctly, and what follows cannot be written off as a coding error.

Now the same thing from 2023 onward:

Both t columns sit below two in absolute terms, and this is exactly where wording has to be careful. We cannot say "the strategy now loses": the minus on the first hypothesis is statistically indistinguishable from zero. There is only one thing we can say, and it is enough - the plus that stood firmly until 2023 is gone.

The paper came out in January 2023.

Two caveats are needed here, without which the picture would look too clean. First: 2026, still incomplete and covering only 31 days, turns positive again. On a sample that size it proves nothing, but leaving it out to protect the story would be dishonest. Second: the clean break belongs to the first hypothesis. The second behaves raggedly across years in both periods - 2021 was already unprofitable before publication, while 2024 and 2025 were profitable after it. It sums to zero across 2023-2025, but its date of death cannot be pinned down.

Coincidence or consequence

Causality cannot be proven strictly on a single pair, and we do not claim it. But the circumstances form a recognisable picture. The authors themselves wrote that the anomaly would disappear once market participants became aware of it, and framed their paper explicitly as a warning. Today, off-the-shelf "Tokyo Fix Gotobi" expert advisors are sold on retail marketplaces as a boxed product. The flow of Japanese payments has not gone anywhere. What changed is how many people know what time it arrives.

The mechanism did not break. What broke was the chance of being on the right side of it before everyone else.

And the inconvenient part

Suppose the anomaly were still alive. Would that mean money?

The answer sits in the same Krohn paper, and it is sobering. The authors state plainly that the returns to the fix reversals, adjusted for transaction costs, are significantly negative for the client. The profit accrues to the side supplying liquidity, not the side taking it. A two basis point effect simply does not cover the spread an ordinary participant pays.

Our own measurements over 2020-2026 confirm this from another angle: the unconditional reversal around the Tokyo fix in USD/JPY is still detectable (roughly one basis point ahead of the fixing), but it is half the size recorded on pre-2019 data, and our costs more than cover it.

Two things are worth carrying away from this when reading any publication with an attractive backtest. A real mechanism and an available trade are different objects: a flow can exist, be measured, and be confirmed by natural experiments, and still leave no room for whoever pays the spread. And a published edge has a shelf life, with the clock starting on the publication date.

This material is for informational purposes only and does not constitute individual investment advice.