Survivor-only stocks flatter a Japanese factor strategy's 0.94 Sharpe

This Japanese strategy probably has a genuine stock-selection signal. Its headline 0.94 Sharpe deserves less confidence: the evaluation period helped select the configuration, and the universe contains stocks already known to have survived.

Mita and Takahashi start with the Fama-French five factors. The textbook model explains today's return using today's factor returns, leaving nothing a trader can act on. The authors forecast both parts of that product. Five LightGBM regressors predict tomorrow's returns for the market, size, value, profitability and investment factors. Their daily macro inputs include Japanese and US 2-year, 10-year and inflation-indexed 10-year yields, USDJPY, USDCNY, a commodity series and the Nomura Bond Performance Index. Each series enters in levels and daily growth rates, alongside three spreads. Each stock's alpha and five betas follow random walks, with a Kalman filter updating the estimates daily.

Multiplying the factor forecasts by the filtered exposures produces expected returns for roughly 750 stocks. The strategy buys the top 40 at equal weights and rebalances at each day's open. The intended edge is to predict which factor will pay tomorrow, then own the stocks most exposed to it. An overlay uses three textbook signals on TOPIX futures: a 5-day minus 25-day regression slope, 25-day 2-sigma Bollinger bands and MACD(12,26,9). The book holds its 40 stocks only when their expected return is positive and those rules signal a rising market. Otherwise, it exits stocks and shorts TOPIX futures.

The data cover 2010 to 2023; out-of-sample testing begins in January 2012. The models refit over a moving 5-year window (minimum 2), with tests in 1-year blocks. Trading costs are 7bp for stocks and 1bp for futures. The reported Sharpe is 0.62 for TOPIX futures buy-and-hold, 0.76 for the long-only portfolio and 0.94 for the combined portfolio. Average return for the combined book is 19.8%, against 12.0% for buy-and-hold. Eight large Japanese active funds reach a maximum net Sharpe of 0.62. The authors deduct an assumed fee equal to those funds' median fee to arrive at the combined strategy's 0.94.

LightGBM supplies the long book's edge over TOPIX

The ablation table tells us the most. Replacing the GBM forecasts with the previous day's factor returns takes long-only Sharpe from 0.76 to 0.57. Average return drops from 16.2% to 12.1%, while drawdown increases from 32.3% to 37.5%. With that naive one-day-lag forecast, the Kalman-ranked portfolio trails the index on Sharpe, 0.57 against 0.62. In this test, the LightGBM forecast accounts for the long book's advantage over TOPIX.

Two findings make that result harder to read. Removing any single macro input puts the combined book's Sharpe between 0.925 and 0.937. Every one exceeds the 0.924 produced by the full input set. Monthly Bayesian tuning with Optuna makes no difference to the combined book's 0.92. No individual feature seems to carry much weight, and beating a one-day lag sets a low bar.

The data handling also looks loose. Bloomberg ticker CL1 identifies the commodity series as WTI front month, whose listed minimum is -37.63. Because the ablation keeps the universe and weights unchanged, equal weighting cannot explain the 0.19 improvement. The paper provides no factor-attribution regression, however, so the forecast could simply be choosing a size or value tilt.

Does the futures short hedge the book?

It acts like a market call. The portfolio switches outright from 40 stocks to short TOPIX futures; there is no intermediate allocation. Volatility changes little: 21.0% for the combined book, compared with 21.4% long-only and 19.5% for the index. Full-period drawdown falls from 32.3% to 27.9%. Since 2019, the reduction is greater, with 20.2% for the combined book against 31.6% long-only. The index's drawdown over that period was 31.5%.

The overlay contributes 0.18 Sharpe across the full period and 0.32 since 2019. The authors highlight COVID-19, when the combined portfolio substantially outperformed. A gain concentrated in episodes such as COVID-19 would rest on a few independent bets, regardless of how many daily observations the sample contains. The authors acknowledge that short positions are relatively infrequent. The indicator parameters are standard settings, to be fair, and the paper describes them as such.

Selection reached into the 2012-2023 test period

Universe membership requires a stock to have been listed on the TSE first section on 01/04/2010 and to belong to the TOPIX 1000 on 12/29/2023. Delisted stocks are excluded. The disclosure is explicit: "Consequently, the results are subject to survivorship bias." In their conclusion, the authors say survivorship "may result in overstated performance" and leave the remedy to future work with a larger dataset. They therefore expect overstatement in figures they still use as headline results.

End-date membership brings future information into the universe as well. A stock that remained in the TOPIX 1000 at end-2023 stayed large, which directly benefits a long-only top-40 portfolio. The long book beats the index by just 0.14 Sharpe. The paper offers no way to estimate how much of that comes from survivors.

Input selection used the evaluation period too. The paper says the strategies "are obtained using the combination of input variables that demonstrated the highest performance during Section 3.1". Dropping the US 2-year yield and Nomura BPI yields the winning 0.942, compared with 0.924 using all inputs. A state-space noise variance of 1 was also selected on investment results, beating 0.1 and 0.01. The authors could reasonably answer that the input gap is only 0.02. They also describe the differences between variance settings as "relatively small." Both replies would be fair. The GBM refits genuinely walk forward, although the surrounding configuration was selected over 2012 to 2023. Survivorship remains the bigger concern.

The significance test stops at zero

The paper says the baseline fails to reject the null in recent subperiods, whereas the proposed strategy delivers "statistically significant performance across all subperiods." That null is a zero Sharpe. Since 2019, daily Sharpe for the combined portfolio is 0.066 with p=0.88%, versus 0.040 and p=8.65% for buy-and-hold. Since 2014, the corresponding figures are 0.048 (p=0.88%) and 0.027 (p=9.55%). An index failing to clear zero over five years mostly tells us about sample length. We did not find a test of the Sharpe difference relative to buy-and-hold, the comparison that matters to a buyer. Even against zero, the long-only portfolio since 2019 has p=5.51%.

The authors report annual turnover around 200%. Charging 7bp implies roughly 14bp a year if 200% measures one side of each trade, or double that if it measures both. There is no breakdown reconciling 200% turnover with daily equal-weight rebalancing and complete switches into futures. The paper says only that shorts are relatively infrequent. The fee is assumed, too, rather than incurred.

We could not rerun the strategy: we lack point-in-time TOPIX 1000 membership and the required JGB, inflation-linked and Nomura BPI series. Porting it to the US would produce a different strategy.

One rerun would change my view. Restore delisted stocks to point-in-time TOPIX 1000 membership and use all inputs. If the long-only portfolio still clears 0.62 by a convincing margin, the forecast is worth building on.