A factor is a measurable trait that sorts stocks: how cheap they are, how profitable, whether insiders are buying. You can use factors two ways. A screen keeps the stocks that pass a set of rules, such as price-to-sales below 1, and drops the rest. A ranking system scores every stock in a universe on several factors and puts them in order. Either way, you’re choosing stocks by traits that have been associated with returns, then testing whether that association holds.
I’ve treated my factor ranking as a main source of systematic, alpha-seeking return in my process, aimed at beating the market after adjusting for risk, though I haven’t been able to back that with evidence. Testing it takes return attribution: splitting results into what the ranking contributed and what my own choices added or cost. That’s part of why I put so much work into backtesting.
I built the process so my portfolio would reflect the factors I ranked on, with the ranking as a floor. I’ve done the vast majority of my buying in its top decile, the top tenth of the list, and rarely gone past its first 100 names in a universe of thousands. When I have, it’s usually been because I was looking for exposure to a theme and developing my theses around it more fully. When that exposure has sat too far down the list, I’ve looked instead for plays on the theme’s second- or third-order effects, in companies it reaches less directly. Those companies are often a more sensible way to get the exposure anyway, because they haven’t yet risen the way a direct play on a consensus theme already has.
In October 2026 I ran simulations of my allocation. Before running them, I suspected I’d hold a concentrated portfolio of stocks from this ranking in my taxable account, since those stocks earn my highest return. Backtests show it for the years before I traded the strategy, and my own results have been consistent with them in the years I’ve traded it. The simulations showed I could best meet my tax and risk objectives in that account without the stock portfolio, because after fees and taxes it wasn’t the asset that best suited my tolerance for drawdowns. My IRAs are another matter. I don’t need that money to smooth my income the way I might need money from my taxable account from time to time, so there I can tolerate a lot of volatility for the return.
The ideas behind factor screens and rankings come from a line of research that started in academic finance and moved into books written for investors.
The academic foundation
Eugene Fama and Kenneth French showed that two simple traits, company size and the ratio of book value to market price, explained much of the difference in average returns across U.S. stocks from 1963 to 1990. Market beta, a measure of how much a stock moves with the market, showed no relation to average returns apart from its link to size (Fama and French 1992, p. 427). Their three-factor model (1993) became a standard benchmark for asking whether a strategy’s returns came from something new or from exposure to known factors. The Nobel committee’s background paper for Fama’s 2013 prize in economics says the approach “has been widely used both as a reference model for academic research and as a practical guide for professional investors” (Nobel committee 2013, p. 40). In 2015 they added profitability and investment, in tests running from 1963 to 2013 (2015, p. 3).
Richard Sloan looked at the quality of earnings, in NYSE and AMEX companies from 1962 to 1991. Earnings made up largely of accruals, rather than cash, persisted less into the next year, and stocks with high accruals went on to underperform. Stock prices acted as if investors looked only at total earnings and didn’t fully account for how much less durable the accrual part was (Sloan 1996, p. 290).
H. Nejat Seyhun studied trades by corporate insiders and found that they carry information. In about 59,000 purchases and sales from 1975 to 1981, insiders tended to buy before their stock beat the market and sell before it lagged, and the rise after purchases was the larger effect. But once the trades were made public, outsiders who copied them couldn’t earn abnormal profits, returns beyond what the stocks’ risk would explain, after trading costs (Seyhun 1986, pp. 189, 196, 210). His book Investment Intelligence from Insider Trading (MIT Press, 1998) extends the work.
The factor zoo
Researchers went on to report hundreds of factors. John Cochrane called it “a zoo of new factors” (Cochrane 2011, p. 1047). Campbell Harvey, Yan Liu, and Heqing Zhu counted more than 300 factors in published and working papers and argued that, given how many had been tried, the usual statistical bar was too low (Harvey, Liu, and Zhu 2016, pp. 5–8). Serhiy Kozak, with Stefan Nagel and Shrihari Santosh, argued that returns across stocks are better described by a few blends of many stock traits than by any short list of single-trait factors, including the popular four- and five-factor models (Kozak, Nagel, and Santosh 2020, p. 271).
Many published factors weaken or fail under stricter tests. Researchers disagree about how many. One large study found that most failed (Hou, Xue, and Zhang 2020, p. 2019); another found that most held up (Jensen, Kelly, and Pedersen 2023, p. 2466). The gap comes from choices in how the tests were run: how much the smallest stocks count, whether returns are adjusted for the market’s overall swings, and how the test allows for the many factors that were tried (pp. 2467–2468).
Published factors have also tended to earn less after publication. Across 97 of them, returns were 26% lower between the end of a study’s sample and its publication, which the authors treat as the most that luck in the original results could explain, and 58% lower after publication. They attribute the difference to investors trading on the published findings (McLean and Pontiff 2016, abstract). Their data covers U.S. stocks and ends in 2013 (pp. 9 and 11). The returns shrank but mostly didn’t vanish: in Jensen, Kelly, and Pedersen’s data, 83% of the factors that were significant in their original papers still had positive returns after the original sample (p. 2495).
From papers to screens and rankings
Marc Gerstein’s Screening the Market (Wiley, 2002) turned this kind of thinking into a practical method for building screens and moving from a screen’s output to a decision. Its four steps are to find stocks worth studying, analyze them, buy the best, and sell the ones that no longer fit. I was reading Marc’s articles and building my methodology on top of his research in the five years before he published that book.
James O’Shaughnessy’s What Works on Wall Street (McGraw-Hill; first edition 1996, fourth edition 2011) tested traits one at a time and in combination over decades of U.S. stock data; the fourth edition reaches back to 1926. Beside each strategy’s return, he reported its base rate: how often it beat the stocks it was drawn from across rolling periods. What I took from it was price-to-sales. I hadn’t considered sales as a measure of value before, and I started building it into my screens. With the fourth edition’s longer history, no single value ratio led in every period, and he combined several into composites instead (AAII Journal, 2019). I hadn’t checked how it has held up until I ran the chat at the end of this guide.
After 2001, I started building ranking systems instead of screens, though for a while I used both. Screeners that compared each stock with its industry and could run backtests cost less than subscriptions that offered ranking systems, so I used screens whenever ranking systems weren’t in my trading budget. For choosing stocks by factors, I consider ranking strictly better than screening.
Richard Tortoriello’s Quantitative Strategies for Achieving Alpha (McGraw-Hill, 2009) grew out of the stock-selection models he built for Standard & Poor’s equity research. He tested traits in seven groups (profitability, valuation, cash flow, growth, capital allocation, price momentum, and risk) on S&P’s point-in-time Compustat data. That data keeps each figure as it was first reported and includes companies that later disappeared, so a backtest can’t use numbers investors didn’t yet have or study only the survivors. He sorted the largest 2,200 stocks into fifths on each trait and counted a trait as useful only if the top fifth clearly beat the average, the bottom fifth clearly lagged it, and the pattern held over time. A ranking that works should put every bucket in order applies the same test to a four-factor ranking system I adapted from his book. Growth on its own didn’t pass, and he advises against screening on any single trait (AAII Journal, 2010). His multi-factor combinations are what I took from the book. Before, I’d run a screen and then look through the ratios of what came back to find the well-rounded stocks. His combinations, which had tested well, gave me ideas for more combinations to backtest in my ranking systems, so the well-rounded stocks rose to the top of the ranks instead.
Christopher Mayer’s 100 Baggers (Laissez Faire Books, 2015) went the other direction. It started from 365 stocks that returned at least $100 for every $1 invested between 1962 and 2014 and asked what they had in common (pp. 7, 9). That’s useful for choosing traits to screen or rank on, though the study looks only at survivors, as Mayer notes himself (p. 9). It shows what the winners shared but not how many losers had the same traits.
Luck can look like skill
Nassim Nicholas Taleb’s Fooled by Randomness (2001) is about mistaking luck for skill. Applied to factor testing, the risk is that you rerun a slightly changed screen or ranking system over the same history until one version looks excellent by chance alone. Rerun a screen enough times and one version will look great by chance covers this problem and what to do about it.
Choosing factors to use
The research supplies candidate traits, each with an economic argument behind it. Before you put one in a screen or ranking system, you should be able to explain why it should work, see that it has held up in more than one period and market, and expect its returns to survive the cost of trading it.
Trace a factor’s research with AI
I hadn’t checked how price-to-sales, the factor I took from O’Shaughnessy, had held up, so I asked Claude, with web search on, to find where it was first published and what later studies found. My first prompt is a template that asks for the factor’s name before anything else, so you can paste it as is and name your own.
October 11, 2026
Me
I want to trace the research history of a stock-selection factor before I rely on it. Don’t start yet. First, ask me for the factor’s name, how I calculate it, and where I came across it. Then search the web and open the sources rather than working from memory, and: 1. Find the earliest published research on this factor or the effect it measures, including practitioner books and journals as well as academic papers. For each early source, give the authors, year, and where it was published; the market and sample period it tested; what it found; and the reason its authors gave for why the effect exists. 2. Find the later research that tested whether the effect held up after it was published, including large replication studies such as McLean and Pontiff (2016), Hou, Xue, and Zhang (2020), and Jensen, Kelly, and Pedersen (2023). For each, say whether the effect weakened, disappeared, or held, over what period, and whether it survived trading costs and held outside the smallest stocks. 3. Put both in one table in date order, with a column for the link. Link each source to the page you opened, such as the paper’s journal page, its SSRN page, or the publisher’s page, not to an article about it. Wherever you name a source in your answer, link it there too. If you couldn’t open a source, list it as not opened and give no link, rather than one from memory. If you’re unsure of a detail such as a year or a sample period, say so rather than fill it in. Don’t tell me whether to use the factor. I’ll decide that after I’ve read the sources.
I want to trace the research history of a stock-selection factor before I rely on it. Don’t start yet. First, ask me for the factor’s name, how I calculate it, and where I came across it. Then search the web and open the sources rather than working from memory, and: 1. Find the earliest published research on this factor or the effect it measures, including practitioner books and journals as well as academic papers. For each early source, give the authors, year, and where it was published; the market and sample period it tested; what it found; and the reason its authors gave for why the effect exists. 2. Find the later research that tested whether the effect held up after it was published, including large replication studies such as McLean and Pontiff (2016), Hou, Xue, and Zhang (2020), and Jensen, Kelly, and Pedersen (2023). For each, say whether the effect weakened, disappeared, or held, over what period, and whether it survived trading costs and held outside the smallest stocks. 3. Put both in one table in date order, with a column for the link. Link each source to the page you opened, such as the paper’s journal page, its SSRN page, or the publisher’s page, not to an article about it. Wherever you name a source in your answer, link it there too. If you couldn’t open a source, list it as not opened and give no link, rather than one from memory. If you’re unsure of a detail such as a year or a sample period, say so rather than fill it in. Don’t tell me whether to use the factor. I’ll decide that after I’ve read the sources.
Claude
Before I start searching, I need three things from you:
- The factor’s name, as you know it.
- How you calculate it: the formula or ranking rule, which fundamentals or prices go in, how often it’s rebalanced, and any universe filters such as a market-cap floor or excluded sectors. If you run it on Portfolio123, pasting the factor or ranking-system definition works.
- Where you came across it: a paper, book, newsletter, forum post, or a P123 ranking system. Even a rough pointer helps, because the source you saw may name or cite the original work.
For the output, I’ll make one table covering both the early research and the later replication studies, with a column linking each source to the page I actually opened. Sources I couldn’t open will be marked “not opened” with no link.
Me
Price-to-sales: a company’s market value divided by its sales over the last 12 months, with lower values ranked higher. I came across it in James O’Shaughnessy’s What Works on Wall Street.
Claude
Below are the sources in one table, in chronological order. For the 1980s–2000s journal articles I could only read the abstract pages, because the full texts are paywalled. Where an abstract doesn’t give the authors’ explanation or sample details, I say so instead of filling them in.
Show the rest of Claude’s answer
| Source | Role | Market & sample | What it found | Early sources: authors’ explanation. Later sources: did it hold, after costs, beyond small stocks? | Link |
|---|---|---|---|---|---|
| Kenneth Fisher, Super Stocks, book, published Sept. 1984 by Dow Jones-Irwin | Early, practitioner book | Not verified | Not verified | Secondhand only. A 1984 Kirkus review says Fisher preferred P/S because he saw profits as an outcome whose calculation can be arbitrary, while revenues are steadier and a better read on a company’s health | Not opened |
| A.J. Senchack & John Martin, 1987, Financial Analysts Journal 43(2), 46–56 | Early, practitioner journal | NYSE and AMEX stocks, 1976–84 | Low-PSR stocks had higher returns and higher risk than high-PSR stocks. They beat an equal-weighted market of similar risk, even among money-losing firms. Low-P/E stocks beat low-PSR stocks on both raw and risk-adjusted terms, and did so in 68% of quarters. Low-PSR firms were even smaller than low-P/E firms. | Not given in the abstract | CFA Institute page |
| Bruce Jacobs & Kenneth Levy, 1988, “On the Value of ‘Value’,” FAJ 44(4), 47–62 | Early, practitioner journal. Sales/price is one of 25 attributes, not the paper’s focus | Over 1,000 stocks, mid-1982 to mid-1987. The abstract doesn’t name the market | In a full multivariate regression with 25 attributes, sales/price was among the attributes producing statistically significant abnormal returns | The authors frame it with evidence that psychology, noise trading and style fads can push prices away from value for long periods | CFA Institute page |
| William Barbee, Sandip Mukherji & Gary Raines, 1996, FAJ 52(2), 56–60 | Early. This is the paper the replication studies credit as the original | 1979–91. The abstract doesn’t name the market | Sales/price and debt/equity explained returns better than book/market or size. S/P captured debt/equity’s role and was the more reliable factor. | Not given in the abstract | CFA Institute page |
| James O’Shaughnessy, What Works on Wall Street, McGraw-Hill, first published Aug. 1996, 4th edition 2011 | Early, practitioner book (your source) | Unsure for the 1st edition. The 4th edition’s publisher description cites data going back to 1926 | Not verified | Not verified | Not opened |
| Makoto Suzuki, 1998, “PSR—an efficient stock-selection tool?,” International Journal of Forecasting 14(2), 245–254 | Early, different market | Tokyo Stock Exchange, 1982 onward, about 14 years | Low-PSR stocks beat low-P/B and low-P/E portfolios six times in 14 years. PSR’s information coefficient exceeded the other measures, including size. | The author breaks PSR into asset turnover and leverage, and ties sales growth to economic recoveries | IAOR abstract record |
| George Leledakis & Ian Davidson, 2001, “Are Two Factors Enough? The U.K. Evidence,” FAJ 57(6), 96–105 | Later, different market | London Stock Exchange. The period isn’t in the abstract | S/P had explanatory power beyond book/market and size, and absorbed debt/equity’s, but did not fully replace book/market and size | Held in the UK in this test. The abstract reports no cost or size-segment results | CFA Institute page |
| R. David McLean & Jeffrey Pontiff, 2016, Journal of Finance 71(1), 5–32 | Later, large replication | US. 97 predictors, sample ending 2013 | S/P is in their set: sales-to-price is named as an example of a valuation predictor. Across all predictors, returns were 26% lower out-of-sample and 58% lower post-publication. Valuation predictors had the smallest post-publication declines, though not significantly different from other types | S/P’s own figures are in the Internet Appendix, which I didn’t open. Returns are not net of costs. Portfolios are mostly equal-weighted extreme quintiles, which tilts them toward small stocks. Post-publication returns were higher for predictors concentrated in costlier-to-arbitrage stocks | Published article PDF, copy on gwern.net (SSRN refused my request) |
| Kewei Hou, Chen Xue & Lu Zhang, 2020, “Replicating Anomalies,” Review of Financial Studies 33(5), 2019–2133 | Later, large replication | US, Jan. 1967–Dec. 2016, financial firms excluded | NYSE breakpoints, value-weighted: the annually sorted sales-to-price spread earned 0.5%/month (t = 2.37); the monthly sorted quarterly version earned 0.59% (t = 2.39) | Held at their single-test hurdle of 1.96, but below the multiple-testing hurdle of 2.78. They use value weights because microcaps are hard to exploit given high transaction costs. In what I read there were no net-of-cost results and no pre/post-publication split for this factor | Authors’ Oct. 2018 version, PDF |
| Theis Ingerslev Jensen, Bryan Kelly & Lasse Heje Pedersen, 2023, “Is There a Replication Crisis in Finance?,” Journal of Finance 78(5), 2465–2518 | Later, large replication | 153 factors across 93 countries; US returns from CRSP beginning 1926 | The authors’ factor file lists “Sales-to-market” from Barbee et al. in the Value cluster. The main text has no result for that factor alone. By theme, 11 of 13 themes replicate at 50% or better, the exceptions being low leverage and size | Across all factors, average alpha was 0.49%/month in-sample and 0.26% after the original sample, a 47% decline. They use non-microcap breakpoints and market caps capped at the NYSE 80th percentile. Replication rates were 77.3% for mega caps and 79.8% for large caps (all factors, not S/P alone). The main text has no trading-cost test | Published version PDF, CBS repository; code repo |
| Andrew Chen & Mihail Velikov, 2023, “Zeroing In on the Expected Returns of Anomalies,” JFQA 58(3), 968–1004 | Later, trading costs | 204 anomalies | After bid-ask spreads, post-publication effects and the post-early-2000s trading era, the average anomaly nets about 4 bps/month and the strongest at best 10 bps | I couldn’t verify that S/P is among the 204 or find its own net return; I read only the abstract | Penn State record |
How the versions compare with your calculation. You use trailing-12-month sales over current market value. Chen and Zimmermann’s open-source documentation describes Barbee et al.’s signal as annual sales over market equity. Hou, Xue, and Zhang report both an annual and a quarterly version. Jensen, Kelly, and Pedersen always use the most recent accounting data, sum quarterly income items over the previous four quarters, and use the latest price. That is the closest match to your definition.
Details I couldn’t settle. For Barbee et al.’s original t-statistic, the Jensen, Kelly, and Pedersen factor file lists 2.21. The Chen and Zimmermann documentation lists about 2.5, from an equal-weighted multivariate regression. I couldn’t check either against the paper. Chen and Zimmermann’s files also map McLean and Pontiff’s “Sales/Price” predictor to the Barbee signal, which confirms it was in that study. Neither the Barbee nor the Jacobs and Levy abstract names the market tested.
Secondhand items. Wikipedia says Fisher’s 2006 book states the PSR is widely known and no longer as useful for finding undervalued stocks. I didn’t open that book.
Not opened. Super Stocks; every edition of What Works on Wall Street; McLean and Pontiff’s Internet Appendix, which has the per-predictor results; the journal version and Internet Appendix of Hou, Xue, and Zhang; the Jensen, Kelly, and Pedersen Internet Appendix and the per-factor data at jkpfactors.com; Chen and Velikov’s full text; and the full texts of the FAJ articles.
Where the sources say something different
Separate AI sessions opened every link in Claude’s answer, which is quoted unchanged, and compared each claim with its source. Every link opened what Claude said it was. The table lists the places where a source says something different, or more.
| Source | What the source says |
|---|---|
| Senchack and Martin (1987) | The abstract says only that the PSR screen “seems to work” on money-losing companies; it doesn’t say they beat the market. The 68% counts raw returns only, not risk-adjusted ones. |
| O’Shaughnessy (1996) | Records disagree on the first edition’s date: a retail listing gives August 1996, the Library of Congress record 1997. |
| Suzuki (1998) | The paper reports sales moving with an index of the business cycle, not causing recoveries, and those data start in fiscal 1970, not 1982. Six wins in 14 years isn’t a consistent win. |
| McLean and Pontiff (2016) | The paper doesn’t say its Internet Appendix reports sales-to-price’s own returns. “Tilts them toward small stocks” is Claude’s inference, not the paper’s. Predictors in costlier-to-trade stocks did earn more after publication, but they earned more before it too; the paper didn’t find that they declined less. The link is an unofficial copy; the publisher’s version is at doi.org/10.1111/jofi.12365. |
| Hou, Xue, and Zhang (2020) | The t-statistic of 2.37 is for one method, NYSE breakpoints with value weights. With equal weights, where small stocks count as much as large ones, the t-statistic is above 4. Value-weighted over the years Barbee and his coauthors studied, the spread fails even the 1.96 bar (t = 1.30). The paper’s first reason for value weights is that they reflect investors’ actual wealth; trading costs come second. These figures are from the October 2018 manuscript Claude linked. |
| Jensen, Kelly, and Pedersen (2023) | The paper does show sales-to-market on its own: Figure 11 plots it among the Value factors worldwide, and its 95% interval sits above zero. The 11-of-13 themes, the 0.49% and 0.26% alphas, and the mega- and large-cap rates are for U.S. factors, not all 153, and the size-group rates count only factors significant in their original papers. Their data assume accounting figures become public four months after a period ends, so the “most recent” sales can be four months old. |
| Chen and Velikov (2023) | The 10 basis points is after controlling for data mining, and the paper adds that combining anomalies nets around 20. Its 204 anomalies come from Chen and Zimmermann’s April 2021 data, which includes sales-to-price, so the factor is almost certainly among them, though the paper doesn’t name it. |
| Chen and Zimmermann | Their file gives Barbee’s t-statistic as 2.52 and rates the signal “likely” rather than “clear.” Their own reproduction earns 0.71% a month (t = 2.86). “Equal-weighted” is their coding of a regression, not necessarily Barbee’s term. |
Claude listed the sources it couldn’t open instead of linking them from memory. Its mistakes were in scope, such as which weighting or which years a result held for, and in one hedge it dropped from an abstract. For price-to-sales, the scope decides the result. The factor looks strong when small stocks count as much as large ones, and in value-weighted portfolios, where large companies count more, it fails over the years its first study covered.
Your check
Run the template on one factor from a screen or ranking system you use. From the sources you’ve opened, write down the economic reason the factor should work, the period it was first tested on, and what later studies found after it was published.