Did portfolios of financially strong companies beat the S&P 500 after costs?
A quality screen helped rank stocks in historical data. The corrected portfolio estimates were positive against the S&P 500 benchmark after costs, but both confidence intervals included zero.
English research archive · Japanese original
A company can be profitable, carry little debt, and rarely issue new shares. Does a portfolio of companies like that beat the S&P 500 after trading costs? In the corrected historical test, the portfolio that weighted the top fifth by market capitalization had estimated annualized advantages of 1.44 and 1.68 percentage points over SPY, an ETF that tracks the S&P 500. Both 95% confidence intervals included zero, so the test did not establish an advantage after costs. The quality score showed a small positive association with stocks’ return ranks over the following year. That ranking result answers a different question from the portfolio comparison.
The U.S. tests covered 2010–2025. Separate historical checks looked at Japanese stocks and two U.S.-listed quality ETFs. The October 2026 research included results from before and after a company identifier correction, labeled separately below.
What went into the quality score
Each week we ranked eligible stocks on three measures and averaged the ranks.
- Net income over the last four quarters divided by equity. Higher ranked better.
- Debt divided by equity. Lower ranked better.
- Growth in shares outstanding from a year earlier. Lower ranked better.
The US figures came from SEC filings available at the time of each decision. Stocks qualified with a share price of at least $10 and average daily dollar volume of at least $20 million over 20 sessions. The universe included companies that later delisted.
1. Did the score help rank stocks one year later?
Before testing, we required a positive 95% interval for rank correlation in both periods and higher returns for the top fifth than for the eligible universe. We measured returns over the next 250 trading sessions.
| Period | Average stocks per week | Rank correlation [95% interval] | Top fifth versus eligible universe, annualized before costs |
|---|---|---|---|
| 2010–2019 | 973 | +0.056 [+0.044, +0.067] | +1.7% |
| 2020–2025 | 1,397 | +0.109 [+0.087, +0.130] | +3.8% |
The table shows the results saved before the company identifier correction. The rerun still met the pre-specified screening criterion. Intervals obtained by resampling longer, one-year blocks to account for overlapping observations also stayed above zero.
At 60, 120 and 250 sessions, correlations were +0.024, +0.034 and +0.056 in the earlier period, and +0.049, +0.073 and +0.109 in the later one. The relationship was stronger over longer horizons in this sample. It gives limited guidance for short trades or individual stocks.
2. Could a portfolio beat SPY after costs?
We specified four portfolios and a separate pass criterion before seeing their results. Each started with $100,000, rebalanced annually, and included commissions and spreads. The table returns are medians across 12 staggered starting months. The pass criterion required the 95% interval for the SPY difference to lie above zero in both periods.
Top fifth, weighted by market capitalization. Results after the company identifier correction.
| Period | Annualized return | Difference versus SPY [95% interval], percentage points |
|---|---|---|
| 2010–2019 | 14.96% | +1.44 [−0.52, +2.94] |
| 2020–2025 | 17.08% | +1.68 [−0.30, +3.20] |
The differences are calculated from annualized daily log-return differences. They can differ from subtracting the displayed annualized returns. The corrected records leave the difference versus the full eligible universe unconfirmed. That comparison had appeared in an earlier version.
Earlier results: equal weights before the identifier correction
| Period | Portfolio | Annualized return | Versus SPY [95% interval], pp | Versus eligible universe [95% interval], pp |
|---|---|---|---|---|
| 2010–2019 | Full universe | 11.06% | −2.00 [−4.06, +0.15] | — |
| 2010–2019 | Top fifth | 13.45% | +0.09 [−2.13, +2.13] | +2.03 [+0.70, +3.52] |
| 2020–2025 | Full universe | 10.28% | −4.19 [−7.59, −0.87] | — |
| 2020–2025 | Top fifth | 13.03% | −1.77 [−6.16, +2.90] | +2.58 [−1.14, +6.35] |
Keep the two versions separate when comparing methods. Within the earlier equal-weight test, the top fifth had a positive estimate against the eligible universe in both periods, and the interval crossed zero in the later one. In that same period the full universe trailed SPY by an estimated 4.19 points. A screen can improve on its starting pool and still fall short of the index.
Two other portfolios in the earlier comparison also missed the adoption criterion. The intervals for an equal-weight portfolio of the top 50 crossed zero in both periods. A version that added sustained sales growth trailed SPY by 3.88 points a year in 2020–2025, with an interval entirely below zero.
A few holdings mattered to the corrected market-cap result. Removing its five biggest contributors in 2020–2025 brought the annualized return down to 7.33%. Its maximum drawdowns were −21.0% and −33.7% in the two periods.
3. Did the screen add anything beyond owning large companies?
A separate review tested the 50 largest stocks drawn from the top quality fifth against the 50 largest eligible stocks overall. Both portfolios started with $100,000, used whole shares, capped each holding at 20% and applied the same costs. Differences are medians of matched starting-month comparisons.
| Period | SPY annualized | Largest 50 annualized | Largest 50 from top quality fifth annualized | Quality version versus largest 50 [95% interval], pp |
|---|---|---|---|---|
| 2010–2019 | 13.26% | 12.89% | 14.64% | +1.86 [−0.60, +4.50] |
| 2020–2025 | 14.93% | 14.83% | 14.43% | −0.45 [−2.82, +2.02] |
The quality version trailed in the later period, so it missed the pre-set requirement to lead in both. Starting with $200,000 left the later difference in the same direction at −1.00 point. The largest 50 without the screen trailed SPY by estimates of 0.37 and 0.09 points, with intervals crossing zero. The results leave the incremental effect of the quality screen uncertain.
Additional evidence: Japan and quality ETFs
For Japan we built the same three measures from 37,830 filings from EDINET, Japan’s corporate disclosure system, covering 4,447 companies. The test ran from July 2017 through September 2025 and covered domestic stocks with a closing price of at least ¥300 and average daily trading value of at least ¥100 million over 20 sessions. It included an average of 1,289 companies per week and measured returns relative to each stock's industry among 33 industry groups.
| Japan rank correlation over 250 sessions | Result |
|---|---|
| Full sample [95% interval] | +0.042 [+0.026, +0.059] |
| 2018–2021 | +0.057 |
| 2022–2025 | +0.020 |
| Small caps | +0.081 |
| Mid caps | +0.012 |
| Large caps | +0.007 |
The Japanese result met its pre-set correlation criterion. It was weaker than the US result and weakened in the later period. In a separate after-cost comparison of 99 monthly starting cohorts, the one-year returns were +10.89% for an equal-weight portfolio of the top fifth, +10.43% for the eligible universe and +14.20% for TOPIX with dividends. The top fifth led the universe in 46 of the 99 cohorts.
Two US-listed quality ETFs trailed SPY by 0.43 and 1.18 points a year over their respective histories of roughly 13 and 21 years through October 8, 2026, using dividend-inclusive prices after fund expenses. This descriptive comparison covered two funds still operating and excluded closed funds.
How much can the tests tell us?
Before testing, we estimated that annual differences of about 4 points in 2010–2019 and 8 points in 2020–2025 would be needed to distinguish an effect. The corrected market-cap estimates were 1.44 and 1.68 points. An effect of that size could be difficult to detect with these samples.
We had examined the 2010–2025 period in other research and developed the longer-horizon question after seeing shorter-horizon results. The US sample covered liquid stocks, the ranking test was before costs, and the portfolio tests included trading costs but excluded taxes. In Japan, the main result excludes about ten companies with erroneous prices after the first run. The original correlation of +0.046 also met the criterion. Another 642 companies lacked price histories. We did not measure how that omission affected the results.
On October 8 we decided to leave this screen out of live trading. When reading a similar backtest, check what the score predicts, what the portfolio earns after costs, how the benchmark was chosen and whether a handful of holdings drive the result.
Updates to the Japanese original
- November 2, 2026. Revised the structure and wording. The research periods and numerical results stayed the same.
- November 6, 2026. Updated the annualized returns, SPY differences and intervals, result excluding the five biggest contributors, and drawdowns for the market-cap portfolio to match the October 6 rerun with corrected company identifiers. An earlier fix removed duplicate company-wide market capitalization assigned to multiple share classes. The 2020–2025 return moved from 18.60% to 17.12% and then to 17.08%. The corrected records left the difference against the eligible universe unconfirmed. The conclusion about the advantage over SPY stayed the same.
AlpAI Lab publishes research and records from its own investing. We do not give personalized investment advice.