This is a standalone summary of an article first published on quantstr.at. The Python and R code, with every equation and regression walked through line by line, is in the original.
TL;DR
- Strategy: the sector rotation of Section 4.1, but funds are ranked by their alpha (the part of their return that the three Fama-French factors do not explain) instead of their past return. From Section 4.2 of Kakushadze and Serur (2018), 151 Trading Strategies.
- Result, January 2000 to September 2026, net of 10 bps trading costs: the top-3 alpha portfolio returned 7.00% a year. Equal weight on the same nine sectors returned 8.70%, the S&P 500 8.24%, and the momentum rotation of Section 4.1 on the same months 8.52%.
- Worst fall from a peak: -47.2% for alpha top 3, against -49.2% for equal weight and -39.2% for momentum.
- The signal is very uncertain: one year of daily returns is a short window for estimating a fund's alpha, and only 1.3% of our monthly alpha estimates were statistically distinguishable from zero.
- It depends on timing: with lagged factor data and next-day execution the same rule earned between 6.9% and 9.1% a year, and no variant beat an equal-weight portfolio that trades the same way.
The idea
The momentum rotation of Section 4.1 ranks sector ETFs by past return. A weakness: a fund can rank first simply because it carries a lot of market risk in a rising market. Alpha tries to fix that. It asks how much a fund earned after accounting for the risks it took. A fund that rose 20% in a year when the market rose 13%, and that moves one and a half times as much as the market, has done nothing special. A fund that rose 20% while the market was flat has.
We apply the rule to the nine original SPDR sector funds: XLB materials, XLE energy, XLF financials, XLI industrials, XLK technology, XLP consumer staples, XLU utilities, XLV health care and XLY consumer discretionary.
The Fama-French model behind alpha
In 1993 Eugene Fama and Kenneth French showed that much of the difference in returns between diversified stock portfolios can be explained by three things:
- Market (MKT): what the whole stock market earned above the Treasury-bill rate.
- Size (SMB, small minus big): small companies have tended to earn more than large ones.
- Value (HML, high minus low): cheap shares (high book value relative to price) have tended to earn more than expensive growth shares.
The three-factor model says a fund's return is explained by how much of each factor it carries, plus a constant. That constant, the regression intercept, is the fund's alpha. A positive estimated alpha means the fund earned more than the model predicts over the estimation period; by itself it is not evidence of a persistent advantage. The factors are published free, daily, in Kenneth French's data library, and that is the data we use.
Methodology
The paper gives one formula, one footnote on how to estimate it and no results. It names no data set, no ETFs and no sample period. So this is one explicit implementation, not a reproduction of published returns.
| Element | Paper, Section 4.2 | This study |
|---|---|---|
| Signal | Alpha, the intercept of Eq. (364) | Same |
| Factors | MKT, SMB, HML | Same, daily file from the Ken French library |
| Returns | In excess of the Treasury-bill rate | Daily ETF return minus the daily risk-free rate |
| Estimation window | "Typically 1 year" | 252 trading days, re-estimated each month |
| Holding period | As in 4.1: 1 to 3 months | 1 month |
| Selection | As in 4.1, for example a top decile | Top 1 of 9 and top 3 of 9, long only |
| Weights, costs | Not specified | Equal weights, 10 bps per unit traded on drift-adjusted turnover |
| Execution | Right after time t | Idealised: the month-end close that produced the signal |
Every result starts in January 2000 and covers 321 months. The strategies and benchmarks use exactly these months, and the momentum rotation of Section 4.1 is rerun on them with the same costs, so it is a like-for-like comparison.
The maths
Each month-end, for each ETF, fit the last 252 days of excess returns to the three factors (the paper's Eq. 364):
R_i is the ETF's daily return minus the Treasury-bill rate, the betas are its loadings on the factors, and the residual ε is what nothing else explains. We fit it by ordinary least squares and multiply the daily alpha by 252 to quote it per year. Then select the N ETFs with the highest alpha and weight them equally (our A1 and A2):
The portfolio trades at the close of month t, earns month t+1, and pays costs on the distance between last month's drifted weights d and the new targets (A3 and A4):
With c = 0.10%. The t+1 means a signal never uses the period it is paid on.
Results
| January 2000 to September 2026 | Alpha top 3 | Alpha top 1 | Momentum top 3 (4.1) | Equal-weight 9 | SPY |
|---|---|---|---|---|---|
| CAGR | 7.00% | 4.48% | 8.52% | 8.70% | 8.24% |
| Annual volatility | 15.86% | 21.30% | 14.75% | 14.69% | 15.13% |
| Sharpe ratio (risk-free = 0) | 0.51 | 0.31 | 0.63 | 0.64 | 0.60 |
| Maximum drawdown | -47.2% | -58.1% | -39.2% | -49.2% | -50.8% |
| Average monthly turnover | 47% | 65% | 46% | 3% | 0% |
Strategies and the equal-weight portfolio are net of the same cost model. Before costs alpha top 3 earned 7.60% a year, so its turnover costs about 0.6 points a year.
What the regressions say
Fitting the model once to the whole sample (6,984 trading days) is descriptive only, because it uses data the early trades could not have had. Still, it is informative. The factors describe the funds sensibly: defensive funds such as XLP, XLU and XLV have low market loadings, financials and technology amplify the market, and the value factor is strongly positive for energy and financials. But none of the nine alphas is statistically different from zero at the usual 5% level. The strongest case is XLF, at -4.2% a year with a t-statistic of -1.96. After accounting for market, size and value, the data cannot show that any sector fund has a lasting edge.
The strategy, though, uses one year, not 28. The regression it ran at the end of September 2026 would have bought XLE (+25.9% a year, t = 1.27), XLK (+18.5% a year, t = 1.38), XLV (+15.7% a year, t = 1.04). Those alphas look large and are very uncertain, with wide intervals. Across all monthly estimates the median absolute t-statistic was 0.58 and only 1.3% exceeded 2 (these one-year windows overlap, so the estimates are not independent). The three funds chosen overlap with the momentum rotation's choice in only 2.0 of three funds on average, so alpha ranking is a genuinely different portfolio, just not a better one here.
What we found
- Alpha top 3 trailed everything in return (7.00% against 8.70% for equal weight and 8.52% for momentum) and beat equal weighting in only 12 of 27 calendar years.
- Concentration hurt more: alpha top 1 returned 4.48% with 21.30% volatility and a -58.1% drawdown.
- Costs: the 47% monthly turnover is the price of rotating on a noisy signal.
Robustness
Window length and number of funds. We re-ran 12 combinations of estimation window (126, 252 and 504 days) and funds held (1 to 4) on one common set of 309 months with fresh capital. CAGR ranged from 4.5% to 9.5%, and 1 of 12 settings beat the equal-weight portfolio's Sharpe ratio on those months. This is an exploration, not 12 confirmations.
Factor availability and execution. The main result uses each day's final factor values at that day's close, but the library publishes them later and can revise them, so a live strategy could not run exactly like this. We tested two changes, alone and together: the regression window ending 5 or 21 trading days before month-end (a stand-in for the publication delay), and trading at the next day's close. Over a common 319 months the six variants earned between 6.9% and 9.1% a year. An equal-weight portfolio traded the same way earned 9.12% (next-day close) and 9.12% (month-end close). No variant beat it, but the gap ranged from about 1.8 points for the main version to 0.1 for the 21-day-lag version traded at the next-day close (9.06% against 9.12%), which also finished above next-day momentum (8.56%). We do not read that as a better signal: it is the same formula on slightly older data, and a swing this large from shifting the data by weeks is consistent with an unstable ranking. This is a robustness check, not a reconstruction of what was actually published on each date.
Limitations
Nine ETFs is a small universe. Execution is idealised (the same close that produced the signal) and costs are a flat 10 bps. The main result assumes final factor values were known at each close, so it is not a fully point-in-time backtest. It is long only, without the paper's optional short leg. Data from Yahoo Finance and the Ken French library can be revised. Past results are not a forecast.
Conclusion
Under the implementation choices used in this study, alpha rotation did not outperform an equally weighted portfolio of the nine original Select Sector SPDR ETFs, and in the main version it did not outperform the corresponding sector momentum strategy over the same period. The universe, weights, rebalancing and costs are our own specification, so this is one historical implementation of the idea in Section 4.2, not a reproduction of the authors' performance. The idea is sound: judge a fund's return after the risks it takes. The difficulty is measurement. One year of daily returns is too short for a precise alpha, so the ranking is highly uncertain and moves with small changes in timing.
Code and full article
The Python and R implementations (written independently and agreeing on every monthly return), the line-by-line walk-through of each equation and regression, and the verification are in the full article at quantstr.at: https://quantstr.at/quant-strategies/alpha-rotation/?utm_source=devto&utm_medium=syndication
If you want the next strategies from the 151 Trading Strategies series, follow @quantstrat here.
Reference: Kakushadze, Z. and Serur, J. A. (2018). 151 Trading Strategies. SSRN: https://ssrn.com/abstract=3247865, Chapter 4, Section 4.2, Eq. (364). 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/alpha...
Here's the R code: quantstr.at/quant-strategies/alpha...