Testing the Persistfolio Methodology Through Time

The purpose of this study was to test a simple question: what would have happened if the Persistfolio methodology had existed years ago and had been applied repeatedly using only the information available at each point in time?

We started with a five-stock portfolio using the five-year Persistfolio methodology. The model was rerun quarterly, just as it would be in practice. At each review, companies were tested using historical data available up to that date, existing holdings were retained while they continued to qualify, and vacancies were filled from the newly ranked qualified companies. Returns were then measured during the following quarter. This is different from taking today's Persistfolio portfolio and backtesting those five stocks. The portfolio had to develop through time based on what the model would have seen at each review.

We ran the study in two stages. The first used today's S&P 500 constituents as a fixed historical universe. This allowed us to test the full walk-forward process, but it introduced an important limitation: the S&P 500 of 2016 was not the same collection of companies that exists today. We therefore ran a second version using historical S&P 500 membership and historical prices, so that the candidate universe itself changed through time.

For both tests we also calculated two versions of the portfolio. The standard Diversified version followed Persistfolio's portfolio construction rules, including its sector limit. The Unconstrained version used the same qualification and ranking methodology but did not apply the sector constraint. This was not the main purpose of the study, but it allowed us to see whether portfolio construction materially affected the result.

Results

The first simulation covered August 2016 through February 2026 using today's S&P 500 constituents as the fixed universe.

These results were strong, but they had to be interpreted cautiously. Using today's constituents historically creates survivorship bias. Companies that disappeared from the S&P 500 were missing, while companies that joined the index later could be considered before they actually belonged to the historical investment universe. The test showed that the Persistfolio process could be run sequentially through historical data, but it did not recreate the universe that would actually have been available to the model.

We therefore repeated the experiment using point-in-time S&P 500 membership. At each historical review date, only companies that belonged to the index at that time were eligible. Companies that later left the index remained part of the earlier universe, while companies that had not yet entered the index were excluded. Historical prices were reconstructed for those constituents, and the same Persistfolio qualification, ranking, retention and replacement process was applied quarter by quarter. The usable period for this test was August 2015 through February 2026.

Correcting the investment universe reduced the historical performance of both Persistfolio portfolios substantially. The result became less spectacular, but the portfolios still finished ahead of the S&P 500.

The most dramatic change, however, was not the overall level of performance. It was what happened to the difference between the two portfolio construction methods. In the fixed-universe simulation, removing the sector constraint increased cumulative return from 658.2% to 1,048.3%. After replacing today's constituents with the actual historical S&P 500 universe, the same comparison became 446.0% versus 451.4%. What initially appeared to be a very large cost of diversification almost completely disappeared.

Discussion

The difference between the fixed and point-in-time universes is fundamental. The fixed-universe test applies Persistfolio historically to companies that are in the S&P 500 today. The point-in-time test instead recreates the companies that were actually in the index at each historical date. The second test is therefore much closer to asking what Persistfolio could actually have selected at the time.

The fixed universe did more than inflate the final returns. It also distorted the apparent value of concentration. Today's S&P 500 is necessarily composed of companies that survived to the present and remained large and successful enough to be included in the index. Applying the Persistfolio ranking retrospectively to that survivor pool gave the unconstrained portfolio an unusually attractive opportunity to concentrate in the highest-ranked historical winners. Once the candidate universe was reconstructed as it actually existed through time, that apparent advantage almost completely disappeared.

This is important because the two portfolio variants use the same underlying research. Persistfolio first determines which companies qualify and ranks them using the same methodology. The difference occurs only during portfolio construction. The Diversified version limits a five-stock portfolio to no more than two companies from the same sector. The Unconstrained version can continue taking the highest-ranked qualifying companies regardless of sector concentration.

In the point-in-time test, the two portfolios were actually identical for the first 18 quarterly periods because the sector constraint did not affect the selections. They diverged only when the highest-ranked qualifying companies became sufficiently concentrated within particular sectors. Even after that divergence, neither construction method consistently dominated the other. By the end of the full period, only 5.4 percentage points separated them.

This suggests that the existing sector constraint did not impose the large performance penalty indicated by the original fixed-universe experiment. In this historical test, Persistfolio obtained most of the same cumulative performance while maintaining greater sector diversification.

The point-in-time result is also more meaningful because the portfolio was not selected once with hindsight and held for ten years. The model repeatedly made decisions through time. Companies qualified and ceased to qualify, holdings were retained or replaced, and the available S&P 500 universe changed. Each quarter's portfolio was determined before the subsequent quarter's return was measured.

Under this test, $1 invested in the Diversified model grew to approximately $5.46, compared with approximately $4.63 for the S&P 500. The Unconstrained model grew to approximately $5.51. These are hypothetical model results, but the comparison is a more demanding test of the methodology than simply examining the historical returns of stocks selected by Persistfolio today.

There are still important limitations. This is a historical simulation rather than live portfolio performance. A five-stock portfolio is highly concentrated, the period represents only one historical market environment, and historical data reconstruction can introduce issues involving corporate actions, ticker changes and data quality. Transaction costs, taxes and individual investor circumstances are also not represented. Most importantly, a methodology that worked during this historical period is not necessarily one that will work during the next one.

Conclusion

The most useful result from this study is not the highest return number.

Our first test produced an extraordinary result, particularly for the Unconstrained portfolio. We then identified an important weakness in that experiment and made the test harder by reconstructing the historical S&P 500 universe. Performance fell substantially. More importantly, an apparent enormous advantage from sector concentration almost completely disappeared. That is exactly the kind of result we wanted the research process to reveal.

After the correction, the five-year Persistfolio methodology still produced five-stock model portfolios that finished ahead of the S&P 500 over the tested period: 446.0% versus 362.6% for the Diversified version and 451.4% versus 362.6% for the Unconstrained version.

One historical simulation cannot establish that the methodology has discovered a persistent source of future excess return. It does, however, give us a reason to continue testing it. The result survived when one of the largest weaknesses in the original experiment—using today's S&P 500 constituents to represent the past—was removed.

The corrected results are less spectacular than the original results. They are also more credible. For this research, that matters more.

Historical and backtested results are hypothetical and are not a projection of future performance. Persistfolio model portfolios are standardized quantitative research models and do not represent actual investor accounts. Results do not include individual taxes, transaction costs or investor-specific circumstances. Persistfolio provides research for educational and informational purposes and does not provide personalized investment advice.