This is a standalone version of an article first published on quantstr.at. The Python and R code, the data files and the line-by-line walk-through are in the original.
TL;DR
- Strategy: Section 4.3 of 151 Trading Strategies ranks funds by the R-squared of a four-factor regression (low R-squared means the fund does something of its own), then by alpha, and buys the low-R-squared, high-alpha group.
- What we did: followed the working-paper design of Amihud and Goyenko (2013), the research the book relies on, on 1,362 US equity mutual funds from Yahoo Finance, 2012 to 2025 (18,677 fund-years).
- Result: the lowest-R-squared, highest-alpha group earned a four-factor alpha of −1.10% a year (t-statistic −4.5). Amihud and Goyenko report +2.81%. The average fund-year alpha here was −1.45%.
- Predictive regression: the coefficient on transformed R-squared is 0.031 (t-statistic 0.52), against −0.680 (t-statistic 7.69) in the original, which also had fund-characteristic controls. We find no significant relationship in this sample.
- Monthly long-short adaptation of the book's rule: 1.37% a year after costs.
- Caveat: survivors-only data and a different period. A failure to reproduce is a finding about this sample, not proof the original was wrong.
The idea
An actively managed fund is meant to differ from the market. Regress its daily returns on market, size, value and momentum factors and look at the R-squared: the share of its movement the factors explain. A fund with an R-squared of 0.98 is close to an index fund with a fee. One with 0.70 is making bets of its own. The book's rule favours low R-squared funds, and among them those with a high alpha, the part of return the factors do not explain. The evidence behind it is Amihud and Goyenko (2013), who studied US equity mutual funds from 1990 to 2007.
Research design
For each fund and calendar year, regress daily excess returns on the four factors and their one-day lags to get alpha, R-squared and the residual standard error. The previous year's values predict the next year's alpha and information ratio (alpha divided by residual standard error). Each year funds are sorted into five R-squared groups and, within each, five alpha groups: 25 groups. The sample keeps funds with at least 125 daily returns in the estimation year and 50 in the test year, and trims the top and bottom 1% of R-squared.
For fund j and year y:
Dy and Ds are year and style dummies, with standard errors clustered by fund.
Data
Yahoo Finance's fund screener returned 5,607 US equity funds in nine style-box categories. 3,156 had price history before mid-2011. We removed 83 with a daily move above 30% (unadjusted splits and similar errors) and merged share classes whose daily returns correlate at 0.999 or more, leaving 1,458 distinct funds, then removed index-named funds: 1,362 funds remain. Factors are Kenneth French's daily four factors. Yahoo lists only funds that exist today, so the sample contains survivors only, and it has no point-in-time fund size, expenses or turnover.
Results: the 25 groups
Average next-year four-factor alpha (% a year, t-statistic in brackets in the original table), rows by previous-year alpha, columns by previous-year R-squared quintile:
| Alpha (y-1) | R2 Low | R2 2 | R2 3 | R2 4 | R2 High | All | Low - High |
|---|---|---|---|---|---|---|---|
| Low | −2.433 (−8.82) | −1.937 (−8.26) | −1.894 (−10.99) | −1.084 (−6.82) | −1.315 (−9.86) | −1.733 (−19.13) | −1.118 (−3.65) |
| 2 | −1.358 (−5.71) | −1.538 (−7.82) | −1.267 (−7.83) | −1.422 (−9.33) | −1.274 (−10.09) | −1.372 (−17.13) | −0.083 (−0.31) |
| 3 | −1.330 (−5.37) | −1.669 (−8.78) | −1.296 (−7.83) | −1.594 (−11.29) | −1.146 (−9.34) | −1.407 (−17.60) | −0.184 (−0.67) |
| 4 | −1.182 (−4.94) | −1.544 (−8.21) | −1.506 (−9.45) | −1.400 (−9.21) | −1.146 (−10.52) | −1.356 (−17.33) | −0.036 (−0.14) |
| High | −1.101 (−4.53) | −1.481 (−8.47) | −1.603 (−9.28) | −1.695 (−11.06) | −1.156 (−9.96) | −1.407 (−17.78) | 0.055 (0.21) |
| All | −1.478 (−13.25) | −1.633 (−18.47) | −1.513 (−20.30) | −1.440 (−21.22) | −1.207 (−22.19) | −1.454 (−39.80) | −0.271 (−2.19) |
| High - Low | 1.332 (3.62) | 0.457 (1.56) | 0.291 (1.19) | −0.611 (−2.77) | 0.159 (0.90) | 0.325 (2.70) |
Every cell is negative. The best cell (low R-squared, high alpha) earned −1.10% (t-statistic −4.5; −1.10% with t-statistic −2.00 across the 14 annual averages). Higher past alpha goes with higher next-year alpha (High minus Low 0.33%, t-statistic 2.70), so there is some persistence. Low-R-squared funds are not ahead of high-R-squared funds overall (−0.27%, t-statistic −2.19). An independent sort gives the best cell −1.23% (t-statistic −6.07).
Results: do R-squared and volatility predict alpha?
The coefficient on TR² is 0.031 (t-statistic 0.52) over the whole sample, 0.092 (t-statistic 1.23) in 2012 to 2018 and −0.132 (t-statistic −1.37) in 2019 to 2025. Total volatility does predict: funds with higher daily standard deviation earned lower alpha (coefficient −1.16, t-statistic −4.28). The information ratio regression gives TR² a coefficient of −0.0057 (t-statistic −7.80), in the direction of the original, but with negative alphas this is partly arithmetic: a large residual risk divides a negative alpha by a bigger number and pushes the ratio toward zero.
Adding Yahoo's current expense ratio, size, turnover and age as proxies (they carry look-ahead and survivorship, so treat this as a sensitivity check) moves the TR² coefficient to −0.134 (t-statistic −2.11). Expenses are strongly negative (coefficient −0.427, t-statistic −4.23): higher-cost funds earned lower alpha.
Year by year
Each year, take the group the strategy recommends (chosen from the prior year) and equal-weight its funds for the calendar year, against all funds and SPY.
The best group beat the average fund in 7 of 14 years, and its alpha was positive in only 5. With funds chosen once a year, $100 became $390 in the best group, $504 across all funds and $696 in SPY (10.21%, 12.24% and 14.86% a year). Fund returns are net of expenses while SPY is a low-cost index fund, in a period when large US stocks led.
For 2026 (January to September, not in the tables above), the best group's alpha was −3.33% a year against −4.43% for all funds, and its nine-month return was 11.68% against 10.12% for all funds and 12.71% for SPY.
Extension: a monthly long-short adaptation
The book describes the rule for ETFs and does not fix the rebalancing frequency. Applying a monthly version to the same funds (252-day window re-estimated each month-end, equal weights, one-month hold, 10 bps cost per unit traded, no borrow fee) lost 1.37% a year after costs and 0.09% before costs, February 2012 to December 2025 (167 months). Open-end mutual funds generally cannot be sold short, so this is a simulation, not a tradeable strategy.
- Estimation window of 126, 252 or 504 days: −1.4% to −0.4% a year; none positive.
- Lagging the factor data by 0, 5 or 21 days and trading at the next close: −2.6% to −1.4%.
- Costs of 0, 10, 25 and 50 bps: −0.1%, −1.4%, −3.2%, −6.3%. A borrow fee of 1%, 2% or 5% gives −2.4%, −3.3% and −6.2%.
- The long-short return has a four-factor alpha of −2.11% a year (t-statistic −1.29). The long leg alone has −1.76% (t-statistic −1.46), against −1.76% (t-statistic −3.57) for all funds.
Limitations
Survivors-only data (every fund still has prices in 2026), no point-in-time fund characteristics, one share class per fund chosen by return correlation, and an annual design that follows the working paper: the published 2013 version reports its main tests on monthly returns, which we did not replicate. The monthly extension was run in Python only.
Conclusion
On 1,362 Yahoo mutual funds from 2012 to 2025, the working-paper design of Amihud and Goyenko does not reproduce their main result: the best group did not earn a positive alpha, R-squared showed no significant relationship with next-year alpha, and the monthly adaptation lost money after costs. A different period or a signal that has since been competed away could explain this; we cannot say which.
Code and full article
The Python and R implementations, the fund list and the results files are described in the original article on quantstr.at: https://quantstr.at/quant-strategies/r-squared/?utm_source=devto&utm_medium=syndication
If you want the next strategies from the 151 Trading Strategies series, follow @quantstrat here.
References: Kakushadze, Z. and Serur, J. A. (2018). 151 Trading Strategies. SSRN: https://ssrn.com/abstract=3247865, Chapter 4, Section 4.3. Amihud, Y. and Goyenko, R. (2013). Mutual fund's R² as predictor of performance. Review of Financial Studies, 26(3), 667-694. Carhart, M. M. (1997). On persistence in mutual fund performance. Journal of Finance, 52(1), 57-82. Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics, 33(1), 3-56.
Educational content, not investment advice. Backtested results are not a forecast of future returns.




Top comments (2)
Here's the Python code: quantstr.at/quant-strategies/r-squ...
Here's the R code: quantstr.at/quant-strategies/r-squ...