USD/JPY Bounced the Day After Every Big S&P Selloff. Then the Yen Broke.

Where the idea came from

This idea did not come from a forum or a research paper. We found it by searching our own data.

We have nine years of 1-minute data on thirteen markets. So we asked a simple question: when one market has a big day, does another market move the next day?

How we searched:

• We looked at ten pairs of markets.
• We tried three sizes of "big day".
• We only used the in-sample years (the years we are allowed to look at while building a strategy).
• One rule mattered more than the statistics: the effect had to go in the same direction in 2017 and 2018 and also in 2019 to 2021.

Most of the results failed that rule. Gold seemed to predict the Australian dollar very well in 2019 to 2021, but before that it predicted nothing. Gold and silver changed direction between the two periods.

One pair passed. After a big S&P 500 down day, USD/JPY went up the next day. This happened in both periods of the data. It was also not just USD/JPY bouncing back from its own move. We checked: USD/JPY's own big days predicted nothing.

Why would this happen?

The reason is old and well known. When stocks fall hard, people buy the yen as a safe haven (a place to keep money safe in bad times). So USD/JPY falls.

The people who buy yen that afternoon are fast traders reacting to the news. The people who sell yen the next day are in Japan:

• retail margin traders. The Bank of Japan has shown that they trade against every rise of the yen.
• life insurance companies and pension funds that buy foreign assets when prices drop.

First comes the quick reaction, then the bigger, slower money. We can name who is on the other side of the trade. That is the first check we do on any idea.

The rule, and what we measured before writing code

The rule:

• At 21:00 UTC, look at the daily return of the S&P CFD.
• If it is below 0.6 standard deviations of its last 60 daily returns (so, a big down day), buy USD/JPY.
• Buy at 23:00 UTC. The hour after the stock market close has few traders, so we wait until the spread is back to about one pip.
• Sell at 20:55 UTC the next day.

We hold for one night. When you hold USD/JPY long overnight, you receive swap (interest paid to you for holding the trade overnight). So this time the cost is actually a credit.

Before any EA (Expert Advisor, a program that trades for you in MT5) existed, two separate scripts measured the rule on 2017 to 2021:

• 215 signals
• +8.9 pips per trade before costs
• all five years were positive
• total costs of 0.4 pips after the swap

So the profit was twenty times bigger than the costs.

We also tested the opposite trade: sell USD/JPY after big S&P up days. It did not make more than its costs. So the strategy only buys. We also found that Friday signals, traded on Monday morning, lost money. So we skip them.

In sample: a profit every year, plus a swap credit

We used real Dukascopy ticks, $3.50 per side commission, and real swap rates. The optimizer found two good areas of settings. Both used the 0.6 sigma (standard deviation) level. The only difference was the stop loss (the price where a losing trade closes):

PF is the profit factor: money won divided by money lost. Max DD is the maximum drawdown, the biggest drop of the account from a high point.

The 40-pip version made these returns in its five in-sample years: +3.3%, +17.5%, +8.2%, +23.9% and +5.6%.

The swap worked as we planned. It paid about 1.6 pips per night, and three times that on Wednesday nights. Over the whole period it was worth $778.

2020 was by far the best year. We wrote this down as a warning before we fixed the settings. A bounce that pays most during a panic depends on panics always looking the same.

Net return by year. In sample the bounce made money every year; out of sample it made money once, in 2024.

Out of sample: the danger we named came true

Out-of-sample means new data that we did not use to build the strategy.

Before the test, we wrote down how the strategy could die: a long period where the yen keeps moving in one direction for days, or the end of the yen as the cheap currency to borrow. 2022 brought both.

• The Bank of Japan kept interest rates low while the US Fed raised them.
• USD/JPY went from 115 to 150.
• Japan's Ministry of Finance stepped into the market to support the yen.
• The daily range of the pair roughly doubled.

A 40-pip stop was rarely hit in 2019. Now it was hit on 101 of the 172 out-of-sample trades. The win rate (the share of winning trades) fell from 59% to 37%.

Out-of-sample equity of the 40-pip version, after costs. The 2024 jump is the August carry unwind; the rest is the new yen market hitting a fixed stop again and again.

Year by year:

• 2022: lost 7.6%
• 2023: lost 1.4%
• 2024: made 19.7%
• 2025: lost 8.6%

The 2024 profit came from one event, the August carry unwind. The S&P fell hard, and USD/JPY jumped back up exactly as our idea said. Everything else was a drawdown.

The swap kept paying, $583 over the period, but it was too small to help. The final result over four years:

• minus $32 on 172 trades
• profit factor 1.00
• maximum drawdown 21%, compared with 4.3% in sample

The strategy failed eight of our checks. The version with the wider stop lost a bit more, with a smaller drawdown. We rejected it.

What we learned

In one way, our search worked as planned. It found a pattern where we could name the other side of the trade, and it held in two separate periods. We also caught and removed the opposite trade and the Friday version before running the tester.

But the search could not see a change in the market that had not happened yet. Five good years is strong proof that a pattern existed. It tells you nothing about whether the conditions that created it will continue.

So our conclusion is narrow. We do not say that bounces after stock market drops do not exist. They did exist for five years, with a fixed 40-pip stop and a one-day hold. But when the pair started moving twice as much, the fixed stop turned the trade into a coin flip, with costs on top.

Maybe a version with a stop that grows with volatility would have survived. But that is a different strategy. It would need its own untouched data for a fair test, and we have no untouched data left for this one.