Gotobi on EUR/JPY: We Changed Nothing and It Still Worked

The most dangerous button in a backtesting program is "optimize". Optimizing means trying many settings to find the ones that did best in the past. Each time you press it, your results fit the past a bit more and the real market a bit less.

So we found a reason to test our Tokyo-fix strategy on a second currency pair. Before we looked at any result, we made a promise: change nothing. Same rules, one test.

It passed.

The idea, in short

Japanese import companies pay their bills on days that can be divided by five. These are the gotobi days. On those mornings their banks buy dollars and sell yen before the Tokyo fix (the official daily exchange rate, set at 9:55 JST). This planned buying pushes the price in a way we can measure in the hours before the fix.

Ito and Yamada wrote about it in their research. We tested it on USD/JPY and got a profit factor of 1.36 on new data. (Profit factor = money won divided by money lost. Above 1 means profit.)

The key point is that this is about the yen. Importers sell yen. If the effect is real, every yen pair should feel it on the same days. EUR/JPY contains the yen, so it should get the same pressure. And we would not need to change a single setting.

That gave us a rare chance. Most backtests answer the question "did these settings fit this market?" A test with fixed rules on a new market answers a better question: "is the effect real?"

The discipline: no settings changed

We copied every setting from the USD/JPY strategy:

• Buy EUR/JPY at 15:55 UTC the evening before a gotobi day. That is nine hours before the fix.
• Close the trade at the fix, 00:55 UTC, in profit or in loss.
• A 60-pip emergency stop loss (the price where we close a losing trade).
• A 3-pip spread filter. We skip the trade if the spread (the cost to trade) is wider than this.
• One trade per gotobi day, risking 1% of the account.
• If the date falls on a weekend, we use the business day before. This is how Japanese payments work.

No optimizer ever ran on EUR/JPY. The pair got exactly two test runs on real tick data. One was an in-sample check (on older data). The other was the final test on new data. That is all.

The in-sample period, 2019 to 2021, showed the effect: 213 trades and a 1.56 profit factor, after commission and swap (the cost or payment for holding a trade overnight):

In-sample check, 2019-2021: $10,000 grows to $11,308 with fixed rules

The one test that counts

Then came the out-of-sample test: 2022 to 2025. These were four years the rules had never seen, on a pair the rules were never built for.

Out-of-sample equity 2022-2026: $10,000 to $12,625 after all costs

The results:

• 283 trades
• Profit factor 1.42
• Win rate 54% (the share of trades that made money)
• Maximum drawdown 6.1% (the biggest drop from a high point)
• Net profit $2,625 on a $10,000 account, risking 1% per trade, after commission and swap

Every year made a profit:

Net profit by year: +$1,317, +$869, +$234, +$205

The same chart also shows the bad news. 2024 and 2025 made much less than 2022. The USD/JPY version showed the same decline. The effect is still there, but it is getting weaker. We say both things because both are true. The chart would show it anyway if we tried to hide one of them.

How much of this was luck?

One equity curve is just one possible order of the trades. So we used the 283 trades to run ten thousand random simulations. Here is what we found:

• The chance that this set of trades ends with a loss: 1.4%.
• The 95% confidence range for net profit is +$331 to +$5,019. So it is still positive even in the unlucky cases.
• When we shuffle the order of the trades, the 95th percentile drawdown is 9.3%. If you want a careful number, use 9.3%, not the real 6.1%.

You can try it yourself:

The question you should ask

"You were buying EUR/JPY from 2022 to 2025. The yen fell a lot in that time. You just followed the trend."

That is a fair question, and it is why we compare against simple benchmarks. We use MAR: yearly return divided by maximum drawdown. Higher is better.

• Buy and hold EUR/JPY over the same period: 8.9% a year, with a 10.9% drawdown. MAR 0.81.
• The strategy: less return, 6.0% a year, but only a 6.1% drawdown. MAR 0.99.

So for the risk it took, the strategy beat holding the pair by 1.22x:

MAR comparison: strategy 0.99, buy-and-hold 0.81, momentum 0.49

The math gives a stronger answer. The strategy is in the market only about 7% of the time. If its profit came just from the yen falling, that time in the market would have earned about 0.65% a year. It earned 6.0%. In the hours before the fix, the pair drifts up about nine times more than its normal average drift. That is what real planned buying should look like. It also matches the in-sample years. Back then the yen went nowhere, and the strategy still made money.

A second question, about swap

A reader asked this about the USD/JPY version, and it matters here too. "You held EUR/JPY overnight and got paid positive swap for it. MetaTrader uses today's swap rate for the whole backtest, not the rate that was true on each night. So part of that profit is interest the trade would not have earned at the time."

The reader is right about how it works. But the effect is smaller on this pair than on USD/JPY. Swap was $263 of the $2,625 net, about 10%. The tester shows swap in its own column for every trade, so we could remove it with simple math and no new test.

With swap set to zero:

• The four years still make $2,362, or +23.6%
• Profit factor 1.37
• Every year still positive: +$1,255, +$807, +$164, +$137
• The share of in-sample profit factor kept on new data goes up a little, from 91% to 93%
• Per lot, the price move earns $36.51 against $3.28 from swap

Swap adds a little to this result. It does not create it. The full explanation is in the USD/JPY post. Swap matters much more there.

What happens next

Next, we will run the strategy on a demo account. This forward test will show whether the weakness we saw in 2024 and 2025 continues. Until then, you can find every trade, every test and the exact settings on the results page.

SimpleMarkets has since tested gotobi independently, on 5-minute Dukascopy and FXCM data from 2010 to 2026, with a random-day test and a check that trades the same hours on every day. Their replication found the effect on EUR/JPY as well.

We changed nothing and it still worked. Whether it keeps working is for the forward test to show.