When the US Market Moves Overnight, How Does Japan React the Next Morning? — Decoding and Reproducing a US-Japan Sector Lead-Lag Strategy¶
For / Key Points
For: Anyone interested in equity investing who has heard that "US market moves affect Japanese stocks the next day" but hasn't seen how to quantify and exploit that link. No math — just the conceptual backbone.
Key Points:
- A strategy that predicts Japanese 17-sector returns from US 11-sector same-day returns was proposed, and "regularized PCA with prior knowledge" dramatically stabilizes predictions
- An independent replication with different data confirmed the algorithm's edge, but the daily-turnover cost structure devours the alpha — live trading is not viable today
- This is a concrete, quantified example showing that a strategy only works when structure, estimation stability, and execution feasibility are all present
What This Strategy Is Trying to Capture¶
The US market (NYSE/NASDAQ) trades during the late night to early morning hours in Japan time. The Tokyo Stock Exchange opens several hours later. During that gap, sector-level moves in the US may not yet be fully reflected in the corresponding Japanese sectors.
For example, if the US technology sector rallies overnight, Tokyo's electrical equipment and precision instrument sectors might "catch up" the following morning — and if this pattern repeats statistically, there is a predictable structure to exploit.
This paper proposes a systematic strategy to harvest that "information delay caused by the time-zone gap."1
Why a Naive Approach Fails¶
The US has 11 sectors; Japan has 17. The straightforward idea would be to estimate 11 x 17 = 187 pairwise relationships — "which US sector going up predicts which Japanese sector going up the next day."
In practice, estimating 187 parameters from roughly 60 days of data (a rolling window) is hopelessly overfit. There are far too many parameters relative to observations, and the model ends up memorizing yesterday's coincidences as if they were laws.
This is where PCA (Principal Component Analysis) comes in.
The Role of PCA¶
The movements of 28 sectors may look chaotic, but a handful of common patterns explain much of the variation. PCA extracts those patterns mathematically.
The three common patterns used in this paper are economically intuitive:
- Whether the overall market is risk-on (buy equities) or risk-off (flee to safe assets)
- Whether the US or Japan is relatively stronger
- Whether money is flowing toward cyclical sectors (steel, banks) or defensive sectors (food, pharma)
Instead of estimating 187 relationships directly, the strategy summarizes US-Japan sector moves along these three axes. It projects the US same-day return onto the three factors, maps the scores to the Japanese side, and generates next-day signals. Compressing the information this way makes the model far less susceptible to noise.
The Weakness of Plain PCA — and What Regularization Fixes¶
PCA is a classical technique, but it has a weakness. When you estimate a correlation matrix of 28 assets from only 60 days of data, the matrix itself is noisy. Applying PCA to a noisy correlation matrix yields common factors that wobble from day to day.
The technical core of this paper is a device called "subspace regularization."
The idea is simple in essence: blend the data-driven correlation matrix with a pre-specified structure (global / country-spread / cyclical-defensive). The paper sets the prior weight at 90% — meaning "10% data, 90% prior knowledge."
That sounds like the data barely matters, but there is a good reason it works. The prior locks in the broad direction; the 10% data contribution captures today's subtle deviations. Letting the data control everything invites overfitting, but fixing the skeleton and delegating only fine-tuning to the data stabilizes the estimates.
Results from the Paper¶
The paper back-tests over roughly 15 years from 2010 to 2025. The benchmarks are simple momentum (buy sectors with recent strength), plain PCA, and the proposed regularized PCA.
| Strategy | Annualized Return | Sharpe Ratio | Max Drawdown |
|---|---|---|---|
| Simple Momentum | 5.63% | 0.53 | -16.97% |
| Plain PCA | 6.24% | 0.62 | -23.65% |
| Regularized PCA | 23.79% | 2.22 | -9.58% |
Regularized PCA dominates. It also has the smallest max drawdown (the largest peak-to-trough decline), meaning the return comes with lower risk.1
Even after Fama-French 3-factor and Carhart 4-factor risk adjustments, annualized alpha exceeds 22% and remains statistically significant. This suggests a return source that cannot be explained by known investment styles such as value or momentum.
Independent Replication¶
Reading to this point, you might wonder whether it really works that well. That same skepticism motivated an independent replication using a different data source.
Test Conditions¶
| Item | Paper | Independent Replication |
|---|---|---|
| US Data | Unknown (likely Bloomberg) | FMP API |
| Japan Data | Unknown | J-Quants Light (JPX-provided) |
| Period | 2010–2025 (15 years) | 2021–2025 (4 years)3 |
| Algorithm | As described | Re-implemented per the paper |
The data source differs and the period is only four years. The goal is not to reproduce the paper's exact numbers but to see whether the same method, run on different data, preserves the structural advantage.
Results¶
| Strategy | Annualized Return | Sharpe Ratio | Max Drawdown |
|---|---|---|---|
| Simple Momentum | -3.93% | -0.47 | -27.27% |
| Plain PCA | +14.28% | 1.41 | -13.73% |
| Regularized PCA | +26.64% | 2.47 | -9.66% |
| Double-Sort | +18.86% | 1.65 | -18.32% |
The ranking of strategies matches the paper exactly. Regularized PCA delivered the highest return and the lowest drawdown, well ahead of plain PCA and momentum. The Sharpe ratio of 2.47 slightly exceeds the paper's 2.22.
A strong 2022 (+49.71%) raises the concern that one good year drives the result, but excluding 2022 still yields an annualized +19.74% with a Sharpe of 1.71.
Robustness Across Parameter Choices¶
Beyond the paper's default settings (3 factors, 60-day window, 90% regularization, 30% quantile), 27 alternative parameter combinations were tested. Every single combination produced a Sharpe ratio above 1.7.
This is an important finding: the result is not an artifact of one carefully chosen parameter set.
Placebo Tests¶
Three placebo tests were run to confirm the algorithm is genuinely generating signal.
| Test | Method | Sharpe Ratio |
|---|---|---|
| Regularized PCA (correct) | Correct signal | 2.24 |
| 1-day lag | Trade on yesterday's signal | 0.60 |
| Random | Randomly assign long/short | 0.34 |
A one-day delay causes a steep drop; random assignment collapses the Sharpe to near zero. The edge depends on signal freshness and the algorithm's selection, which is supporting evidence for a genuine predictive structure.2
Why "Profitable" Is Still Too Strong a Claim¶
At this point you might be eager to trade. But treating the paper's numbers as an expected live PnL would be a mistake.
This strategy buys the top 30% of 17 sectors, shorts the bottom 30%, and closes everything at the end of the day — every single day. Daily full-portfolio turnover means a very high number of trades.
Commission can be zero in Japan (certain flat-fee brokerage plans for trades under JPY 1 million/day). The real problem is slippage.
Slippage is the gap between the price assumed in the backtest and the price actually filled. In thinly traded ETFs, your own order moves the price. Unlike commissions, this structural cost cannot be eliminated.
Cost Sensitivity Analysis¶
Here is what happens as one-way slippage increases.2
| Slippage (one-way) | After-Tax Ann. Return | Sharpe Ratio | Max Drawdown |
|---|---|---|---|
| 0 (ideal) | +21.31% | 2.14 | -10.06% |
| 0.01% (1 bps) | +10.90% | 1.20 | -17.47% |
| 0.02% (2 bps) | +1.88% | 0.27 | -25.98% |
| 0.03% (3 bps) | -7.47% | -0.66 | -38.92% |
Between 0.01% and 0.02% one-way slippage, the Sharpe ratio crashes from 1.20 to 0.27. This nonlinear collapse reveals how thin the strategy's gross margin really is.
Why Is It So Cost-Sensitive?¶
Look at the trading mechanics: the strategy holds 5 long and 5 short positions daily, with a 62% daily name-turnover rate. That adds up to roughly 3,000 round-trip trades per year.
Even a 0.01% one-way slip, compounded over 3,000 trades, amounts to roughly 10% annualized drag. With a 26% gross edge, subtracting 10% in costs and 5% in taxes leaves about 11%. At 0.02% slip, the annual cost reaches 20% — nearly wiping out the edge entirely.
The Liquidity Wall and Dead-End Detours¶
TOPIX-17 sector ETFs vary enormously in liquidity. Only 4 of 17 have daily turnover above JPY 100 million; 10 trade below JPY 50 million.
Restricting to liquid names was tested, but narrowing from 17 to 7 ETFs caused the strategy to collapse completely. With only 2 long and 2 short positions, diversification is too thin and individual-name noise drowns the signal.
Dropping the short leg and going long-only was also tested, but most of the edge resides on the short side, and the strategy turns negative.
The map is accurate, but driving the route costs more in fuel than the destination is worth. And every detour — fewer names, no shorts — is a dead end.
Lessons from This Paper¶
In investment strategy research, "how to stabilize estimation" matters as much as "what to predict." This paper demonstrates that clearly.
Plain PCA and regularized PCA use the same information (the correlation matrix of 28 US-Japan sectors) and the same prediction structure (low-rank linear forecasting). The only difference is estimation stability. That alone lifts the Sharpe ratio from 0.62 to 2.22.
At the same time, no algorithm — however clever — can succeed if the edge does not exceed execution costs. This replication is a quantified example that a strategy only becomes viable when structure, estimation stability, and execution feasibility are all present.
"Interesting research" and "profitable trading" are different things. The ability to see where the boundary lies is, in quantitative investing, as valuable as the strategy itself.
Nakagawa K., Takemoto Y., Kubo K., Kato M. (2026) "Investment Strategy Exploiting US-Japan Sector Lead-Lag Relationship Using Subspace-Regularized PCA," JSAI SIG Technical Reports, SIG-FIN-036-13, pp.76-83. https://doi.org/10.11517/jsaisigtwo.2026.FIN-036_76 ↩↩
Independent replication. Data from FMP API (US) and J-Quants Light (Japan) for 2021–2025; the paper's method was re-implemented and sensitivity analysis was conducted under varying cost assumptions. ↩↩
The J-Quants Light plan only provides TOPIX-17 sector ETF data from March 2021 onward, so the paper's full period (2010–2025) could not be replicated. ↩