If You Purify Dead Momentum, Will It Come Back to Life? — A Japanese Stock Replication of Residual Momentum (2023–2026)

*Verification Date: July 17, 2026 / Author: Kettle*

Summary

The momentum effect, which suggests that "stocks that have gone up will continue to go up," has been confirmed in markets around the world, but Japanese stocks are the most famous exception. There is a standard prescription for this in academia: "residual momentum," which measures momentum using only the stock-specific portion (residuals) after removing the parts of returns that can be explained by factors such as the market or value. Following the original research from the U.S., there have been multiple reports stating that "it works in Japan if you use residuals."

In this article, I used daily residuals estimated by my own Barra-style risk model to replicate this standard approach using the most recent 35 months of J-Quants data. The conclusions are the following three points:

1. **It did not come back to life.** The predictive power (IC) of stock-specific momentum was actually slightly negative (−0.029, t=−1.94), and the series that completely extracted only the idiosyncratic portion had an annualized return of −3.2%.
1. As a byproduct, a methodological lesson was learned: **Even if you residualize the scores, the portfolio is not residualized.** In the long-short portfolio constructed with residual scores, 72% of the risk was still the momentum factor.
1. The source of the profit from raw momentum during this period (annualized +8.4%, no costs) was not stock-specific momentum, but rather **momentum on the factor side** — this is my current interpretation (hypothesis).

Note that all long-short (LS) portfolios in this article are **analytical tools** for measuring factor performance and are not intended for actual trading. Please see the disclaimer at the end of the article for details.

-----

1. Japan is the "Graveyard of Momentum"

The momentum effect — the phenomenon where winners over the past 6 to 12 months continue to win thereafter — is one of the most robust anomalies in empirical finance, confirmed across countries and asset classes since Jegadeesh and Titman (1993).

However, Japanese stocks are a globally famous exception. Asness (2011), in a paper titled "Momentum in Japan: The Exception That Proves the Rule," confirmed that the Sharpe ratio of momentum alone in Japan is nearly zero (the main point of the paper is a defense that if you view both as a "system" based on their strong negative correlation with value, it works even in Japan).

It was the same in my verification environment. In the factor verification for this series, the IC (rank correlation between scores and the following month's returns) of raw 12-1 momentum (12-month returns excluding the most recent month) was statistically indistinguishable from zero.
https://note.com/kettle_3/n/n47c06e9ad9a2
In this verification window (August 2023 to June 2026, 35 months), the IC of raw momentum was +0.007 (t=+0.35). The world-standard anomaly is quietly dead in Japan.

2. The Standard Prescription: "Residual Momentum"

The standard academic prescription for this is **residual momentum (or idiosyncratic momentum)**.

Blitz, Huij, and Martens (2011) used U.S. data to decompose stock returns into "parts explainable by factors" and "residuals" via time-series regression on the Fama-French 3-factor model. They showed that constructing a portfolio using 12-1 momentum based only on residuals (standardized by residual volatility) resulted in risk-adjusted returns approximately twice as high as normal momentum, with smaller fluctuations over time. The logic is as follows: normal momentum strategies implicitly accumulate tilts toward factors that happened to rise over the past year (such as the market or small-cap stocks). By measuring with residuals, you can remove those "fake winners" and extract only the stock-specific momentum.

And what is important for Japan is the follow-up research. Chaves (2016) confirmed this effect in the U.S. plus 21 other countries and reported that it holds true even in Japan, where traditional momentum does not work (the title during the working paper stage was "Eureka! A Momentum Strategy that Also Works in Japan"). Chang, Ko, Nakano, and Rhee (2018) verified the effectiveness of residual momentum specifically for the Japanese market, and Blitz, Hanauer, and Vidojevic (2020) also explicitly stated that it "works in Japan" in their international verification spanning developed and emerging countries.

In other words, the implication of previous research is: **"Dead Japanese momentum comes back to life if you purify it."**

I have a Barra-style risk model built for this series, and the daily residuals (ε: stock-specific returns that cannot be explained by factors) for all stocks have already been estimated. https://note.com/kettle_3/n/n448c84ac7d15
The conditions for replication are in place.

3. Verification Design

**Risk Model**: A simple Barra-style characteristic model. Market + 5 styles (beta, size, value, dividend yield, momentum), no industry factors. Pure factor returns and residuals ε are estimated using daily cross-sectional weighted least squares (WLS). The correlation between the market factor and TOPIX is 0.998, and the basic operation of the model is sound.

**Universe**: Approximately 1,500 to 2,000 stocks (monthly median of about 1,800) that passed the liquidity filter (daily trading value of 50 million yen).

**Score**: At the end of each month, the 12-1 cumulative of ε (idiosyncratic returns over the past 12 months excluding the most recent month). I prepared two types: a raw score version and an adjusted version (Blitz style) divided by the stock's own residual volatility. The comparison target is raw 12-1 momentum for the same period and universe.

**Portfolio**: Equal-weighted LS with top 20% long and bottom 20% short based on score quintiles. Monthly rebalancing, no trading costs.

**Period**: August 2023 to June 2026, 35 formations (constrained by the history length of ε).

There are three main differences from previous research that I will state in advance. (1) The sample is 35 months, which is orders of magnitude shorter (previous research covers decades). (2) The residualization method is different (the mainstream in previous research is time-series FF regression of monthly returns; this verification uses a daily cross-sectional characteristic model). (3) Since it does not include industry factors, industry effects remain in ε. Therefore, this verification is not a "reproduction experiment under identical conditions" but a **replication using a different implementation of the same philosophy**.

4. Results (1): It Did Not Come Back to Life

| Series | IC Average (n=35) | t-value | LS Annualized (No Cost) | Sharpe Ratio |
|--------------|----------|-----|-----------|-------|
| Raw Momentum (12-1) | +0.007 | +0.35 | +8.4% | 0.69 |
| Residual Momentum (Raw Score) | −0.029 | −1.94 | — | — |
| Residual Momentum (Vol Adjusted) | −0.024 | −1.59 | +1.6% | 0.18 |
| Fully Purified w’ε (See text) | — | — | −3.2% | −0.77 |

*IC is the monthly average of Spearman rank correlation. For LS, only the volatility-adjusted version is aggregated for the residual version.

The IC of idiosyncratic momentum was −0.029 (t=−1.94) for the raw score version and −0.024 (t=−1.59) for the volatility-adjusted version. **The sign is actually negative** — winners of idiosyncratic returns are slightly more likely to lose the following month. Both implementations have negative signs, which is internally consistent. However, t=−1.94 barely misses the 5% two-sided significance level, so with a 35-month sample, all that can be said is that it "suggests a weak contrarian tendency."

The LS performance is also consistent. Raw momentum is +8.4% annualized (Sharpe ratio 0.69), while the residual version is +1.6% (0.18). And the series that identity-wise subtracted the factor portion from the portfolio returns to extract only the idiosyncratic portion (w’ε: synthesized residual ε using portfolio weights w) was **−3.2% annualized** (−0.77). The more you purify it, the lower the performance becomes.

[Figure 1]

Figure 1: Left shows the cumulative returns of the 3 series (monthly, no costs, 35 months). Right shows the average variance share of raw/residual momentum LS. Even in the residual version, 72% of the variance is the momentum factor, and stock-specific variance is only 7%. (Created by the author based on Barra-style simple model/J-Quants API)

5. Results (2): Purifying Scores is Not Purifying Portfolios

Perhaps the most valuable takeaway from this verification is this discovery.

When measuring the factor exposure (standardized basis) of the LS constructed with residual scores, the momentum exposure is +2.27. It has hardly decreased from the +2.60 of the raw momentum LS. Looking at the breakdown of variance (risk), 72% of the residual LS is derived from the momentum factor, and the stock-specific portion is only 7% (Figure 1, right). **Even though factors were removed from the scores, the factors remained almost entirely in the portfolio.**

Why is this? When measuring the rank correlation between winners of ε and winners of total returns, it is 0.870 — they were almost the same lineup. The secret lies in the model's explanatory power. The daily cross-sectional R² is 0.10 on average, meaning about 90% of the return differences between stocks are originally ε (idiosyncratic portion). Even if you subtract the factor portion, the rankings of the stocks hardly move, and "residual winners" are almost the same as "just winners." And if you buy those winners, the characteristic of 12-month returns — that is, momentum exposure — automatically increases.

Let's also confirm the numerical consistency. The +1.6% annualized return of the residual LS can be broken down by identity (LS return = exposure × factor return + idiosyncratic portion) into "factor portion approximately +4.8%, idiosyncratic portion −3.2%." All positive performance is from the factor side, and the pure idiosyncratic portion is negative — which is consistent with the negative sign of the IC.

(Technical note: A static approximation of average exposure × average factor return only yields +1.3%, which is significantly lower than the actual factor portion of +4.8%. The difference is the co-variation of exposure and factor returns, or a timing component. One of the lessons learned this time is that static approximations should only be used as a rough guide.)

6. Interpretation: The Source of Profit is Factor Momentum (Hypothesis)

Reading the above straightforwardly, it comes to this:

**In Japan during this period, the source of momentum profit lies in the momentum of factor levels, and there is no momentum at the stock-specific level (in fact, there is weak mean reversion).**

In fact, the pure momentum factor of this model itself rose at an annualized rate of +3.5% (IR 0.74) during this period, and the +8.4% of the raw momentum LS can be largely explained by "high momentum exposure × rising momentum factor."

This resonates with recent research. Ehsani and Linnainmaa (2022), focusing on U.S. data, argued that the profit of individual stock momentum is the accumulation of the autocorrelation of each factor (factor momentum), and that momentum is not an independent risk factor but a strategy for timing factors. The shape observed in this Japanese cross-section — momentum on the factor side, but not on the idiosyncratic side — is consistent with this view. However, this is an interpretation based on 35 months of observation and remains a hypothesis.

7. Why is it the opposite of previous research?

In contrast to multiple previous studies stating that "residual momentum works in Japan," this verification yielded the opposite result. I will honestly list the candidates.

First is the **period**. Previous research covers decades, while this verification covers 35 months. Even if residual momentum had a long-term premium, there is no guarantee it can be observed in a window of less than 3 years (conversely, the possibility that it stopped working after the research was published cannot be distinguished from this sample).

Second is the **residualization method**. The mainstream in previous research is time-series FF regression of monthly returns (36-month window), but this verification is a daily cross-sectional characteristic model. If the definition of residuals differs, the content of "residual momentum" can also change. As seen in Section 5, in this implementation, residualizing scores did not lead to the purification of the portfolio, which is a consequence of the method selection.

Third is the **details of the design**. The lack of industry factors (industry effects remain in ε), the equal-weighting, and the 50 million yen trading value filter universe (where small-cap stocks are relatively thick) are also factors that could affect the results.

Therefore, the position of this article is not a denial of previous research, but a **negative replication** stating that "it did not reproduce with this design and this period." Replication is strong when it succeeds, and even if it doesn't, recording it leads to the next step — this is the policy of this series.

8. Summary and Next Steps

- When the standard prescription "residual momentum" was replicated in Japanese stocks over the last 35 months, the IC of idiosyncratic momentum was a small negative (−0.029, t=−1.94), and the completely purified idiosyncratic portion was −3.2% annualized. **It did not come back to life.**
- **Purifying scores is not purifying portfolios.** Even in the LS of residual scores, 72% of the variance remained the momentum factor (rank correlation of winners 0.870, momentum exposure +2.27).
- The interpretation (hypothesis) is that the source of momentum profit during this period lies in the momentum on the factor side. This is also consistent with factor momentum research.

The next step is also clear. If you really want to extract only idiosyncratic momentum, you need to construct a **portfolio that binds factor exposure to zero at the construction stage** (LS with constraints such as momentum exposure neutrality) rather than tweaking the scores. I will announce this as a future verification task.

-----

Disclaimer

- This article was created for information and research purposes and is not intended as investment advice or solicitation based on the Financial Instruments and Exchange Act. It does not recommend the buying or selling of specific stocks or financial products.
- The long-short (LS) portfolios in this article are analytical tools for measuring factor performance and are not intended for trading. They include short selling, and the author does not operate these portfolios.
- All verification results are observations based on past data and do not guarantee or imply future results. Please note that the sample period is short at 35 months, and statistical uncertainty is high.
- All returns mentioned do not take into account trading costs, taxes, etc.
- While care has been taken in the acquisition and processing of data, the accuracy and completeness of the content are not guaranteed. Please make any decisions based on the information in this article at your own responsibility.

Data Sources and Charts

- Data: Created by the author based on J-Quants API (JPX Market Innovation & Research, Inc.).
- All charts, analyses, and calculations in the article are by the author (Kettle).

References

- Jegadeesh, N., & Titman, S. (1993). Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency. *Journal of Finance*, 48(1), 65–91.
- Asness, C. S. (2011). Momentum in Japan: The Exception That Proves the Rule. *The Journal of Portfolio Management*, 37(4), 67–75.
- Blitz, D., Huij, J., & Martens, M. (2011). Residual Momentum. *Journal of Empirical Finance*, 18(3), 506–521.
- Chaves, D. B. (2016). Idiosyncratic Momentum: U.S. and International Evidence. *The Journal of Investing*, 25(2), 64–76.
- Chang, R. P., Ko, K.-C., Nakano, S., & Rhee, S. G. (2018). Residual Momentum in Japan. *Journal of Empirical Finance*, 45, 283–299.
- Blitz, D., Hanauer, M. X., & Vidojevic, M. (2020). The Idiosyncratic Momentum Anomaly. *International Review of Economics and Finance*, 69, 932–957.
- Ehsani, S., & Linnainmaa, J. T. (2022). Factor Momentum and the Momentum Factor. *Journal of Finance*, 77(3), 1877–1919.