How Do Rising Japanese Interest Rates Affect Stocks? — Testing EWJ with Data

---

Japanese interest rates have started moving again.

In investment commentary, rising rates are often described as a tailwind for banks and a headwind for growth stocks and real estate. The logic is easy to understand, but one question remains:

How clearly does that relationship actually appear in the data?

Stock prices are influenced by more than interest rates alone. Currency moves, overseas equities, oil prices, economic conditions, corporate earnings, and many other factors can all move at the same time.

So in this analysis, we included Japanese long-term interest rates alongside multiple other factors and examined their relationship with Japanese equity performance over the following 60 days.

The target asset in this analysis is the iShares MSCI Japan ETF (EWJ).

EWJ is a U.S.-listed ETF designed to track the MSCI Japan Index, which consists of Japanese equities. Because EWJ is traded in U.S. dollars, it is somewhat different from analyzing a yen-denominated Japanese equity index such as TOPIX.

In other words, this analysis is not looking at Japanese equities purely in yen terms. It is closer to the experience of an overseas investor holding Japanese equities in U.S. dollars. Currency movements between the yen and the U.S. dollar therefore also affect EWJ’s returns.

We used EWJ’s 60-day forward return as the target and analyzed it together with multiple factors, including Japanese long-term interest rates, currencies, overseas equities, oil prices, and macroeconomic data.

Because some of the factors have limited historical availability, the final analysis period is broadly restricted to data from around 2015 onward. We will return to this limitation later.

One result immediately stood out: Japanese long-term interest rates.

The 12-month change in Japanese long-term rates ranked second in model importance among the factors included in this analysis. The 3-month change also ranked relatively high.

This does not mean that interest rates were the second-most important force driving Japanese stocks.

Here, model importance refers to how strongly each factor influenced the model relative to the other factors included in the analysis, measured using SHAP.

A factor can have high model importance without telling us whether a higher value should be interpreted as positive or negative for future stock returns.

Still, it was interesting that changes in Japanese long-term rates carried substantial information within this particular group of factors.

Another important point is that there is no single variable called simply “interest rates.”

We could look at the current level of interest rates, the change over the past three months, or the change over the past 12 months. Even though all of them are based on the same Japanese long-term rate, they represent different factors.

In this analysis, the 12-month and 3-month changes produced different patterns.

Asking only whether “Japanese interest rates affect stocks” is probably too broad a question.

Which interest rate are we using? Are we looking at the level or the change? If we are looking at the change, over what period?

Only after defining those details can we properly test the relationship with data.

Next, we looked more closely at the 12-month change in Japanese long-term interest rates.

We divided the factor into ten groups from lowest to highest and calculated the average subsequent 60-day return for EWJ in each group.

In this sample, larger increases in Japanese long-term rates over the previous 12 months tended to coincide with relatively stronger subsequent 60-day performance.

That is somewhat surprising if we begin with the simple idea that rising rates are automatically negative for stocks.

But this chart does not allow us to conclude that:

“Rising Japanese interest rates cause Japanese stocks to rise.”

For example, stronger economic growth or inflation could have pushed interest rates higher while corporate profits and stock prices were also rising. Currency moves or overseas equity performance may also have been moving at the same time.

Observing a relationship between rising rates and subsequent returns is different from showing that rising rates caused those returns.

The shorter-term change in rates produced a different result.

The 3-month change in Japanese long-term rates also showed relatively high importance within the model. But when we divided the observations into deciles, the relationship was not a clean one in which larger rate increases consistently led to higher 60-day returns.

Some of the lower deciles also showed relatively strong returns, while several middle-to-higher deciles were weaker. The highest decile turned positive again.

That is an important result.

A factor being important to the model is not the same as saying that higher values of that factor are always positive for stocks.

Even for the same Japanese long-term interest rate, the 12-month and 3-month changes produced different patterns.

We show the 12-month result because it is easier to interpret visually, but the 3-month result makes it even clearer that the relationship cannot simply be reduced to “rising rates = higher stock prices.”

This distinction is especially important when working with investment data.

A factor may have high model importance without implying that:

“when this factor rises, stock prices will also rise.”

Likewise, higher forward returns observed under certain conditions do not, by themselves, establish causality.

In this analysis, we therefore do not look at model importance alone. We also examine the relationship with future returns, decile-level performance, and the relationship between factor values and SHAP values.

SHAP is useful for understanding how the model used each factor. But SHAP itself does not prove what caused the market to move.

Importance, direction, correlation, and causality are different concepts and should be treated separately.

One of the clearest lessons from this analysis is how difficult it is to explain markets with only one factor.

When Japanese interest rates move, the rest of the market does not stand still. USD/JPY may be moving at the same time. U.S. equities may be rising or falling. Oil prices may be changing. Growth, inflation, and corporate earnings may also be shifting.

This makes statements such as:

“Rates went up, so stocks went up.”

easy to understand, but potentially too simplistic.

What we really want to examine is how much information each factor carried within a market environment in which many factors were moving simultaneously.

The fact that the 12-month change in Japanese long-term rates ranked second in model importance should be interpreted in that context.

It does not mean that investors should focus only on interest rates.

A more appropriate interpretation is:

Within the particular group of factors used in this analysis, changes in Japanese long-term rates contained information that the model considered meaningful.

There is another major limitation to this analysis.

Because the available history of the selected factors differs, the model is based primarily on data from around 2015 onward. FactDecode does not artificially fill periods before data existed. The model is built using periods in which the required factor data was actually available.

Japan also spent a very long time in an extremely low-rate environment.

That means we do not have a large number of historical observations resembling the current environment in which Japanese interest rates are moving more substantially.

This is one of the difficult parts of quantitative analysis.

A weak relationship in historical data is not the same as evidence that the relationship does not exist.

The opposite is also true. Even if we observe a relationship in the current sample, we cannot assume that it will necessarily repeat during future rising-rate regimes.

If Japan is moving into a different interest-rate regime, we need to keep observing the new data as it accumulates rather than fixing our conclusions based only on the past.

“Higher interest rates should benefit bank stocks.”

“Rising rates should be a headwind for growth stocks.”

There is nothing wrong with forming hypotheses like these.

The problem begins when we treat a plausible economic explanation as if it were already confirmed by the data.

If a hypothesis is going to be used in an investment decision, we want to know how returns behaved under those conditions in the past, whether the factor still carried information when combined with other variables, and how outcomes changed across different factor levels.

At a minimum, those are the kinds of questions we want to test first.

A hypothesis is the starting point of analysis.

Form the hypothesis.
But until it has been tested and supported with data, treat it as a hypothesis.

The relationship between Japanese interest rates and EWJ is a good example of why that distinction matters.

In this analysis of EWJ’s 60-day forward returns, the 12-month change in Japanese long-term interest rates ranked second in model importance among the factors we included.

The 12-month change showed relatively stronger subsequent returns on the higher-rate-change side, while the 3-month change did not show the same simple pattern.

The conclusion is therefore neither “rising rates = higher stocks” nor “rising rates = lower stocks.”

The result depends in part on how the interest-rate factor is defined. And in real markets, currencies, overseas equities, oil prices, economic conditions, and many other factors are moving at the same time.

Instead of explaining the market through a single variable, it is more useful to examine multiple factors together and ask which ones carried information within that broader environment.

And when the answer is not simple, we should resist the temptation to make it simple.

That is an important part of using data in investment analysis.

FactDecode is the analysis environment we are developing to examine assets using multiple factors, including model importance, correlation with future returns, decile performance, and SHAP-based model behavior.

→ Start analyzing with FactDecode