Working paper, June 2026

Profitability, relative to peers.

How profitable is a firm? The answer depends on the comparison. The standard factor model compares every firm with the whole market. I compare each firm with its closest peers, identified from the text of regulatory filings, and find that investors are paid for the difference.

+4.57%Five-factor alpha per year, t = 5.00
0.63 → 0.88Out-of-sample Sharpe, added to five factors
354Months tested, 1996 to 2025

I thank Professor Eugene F. Fama for his valuable feedback over the Spring 2026 quarter of the Research Projects: Finance seminar at the University of Chicago, and the other students in the seminar for their comments.

The question behind the paper

Compared to whom?

↓→Scroll or use the arrow keys. Each step adds one piece of the argument.

01 / The question

The same margin, two different businesses.

Fig 1

The example opens the paper. The peer margins shown here are illustrative.

  1. Two companies report the same 4% profit margin.One is a grocer and the other is a software company.
  2. Among grocers, a 4% margin is a strong result.Grocers typically earn about 2%, so this firm is ahead of its industry.
  3. Among software companies, the same 4% is a poor one.Software companies typically earn about 22%, which leaves this firm far behind its industry.
  4. So what does a 4% margin say about a firm?Very little, until you know who the firm is being compared with. The question I ask in this paper is which comparison group investors should use.
02 / The model

How returns are explained today.

Fig 2
MKTStocks − bills

Market

The stock market's return over Treasury bills.

SMBSmall − big

Size

Small firms minus large firms.

HMLCheap − expensive

Value

High book-to-market minus low.

RMWProfitable − unprofitable

Profitability

Robust minus weak operating profitability.

CMAConservative − aggressive

Investment

Low asset growth minus high.

Fig 3
Explained by the five factorsUnexplained

A schematic. I show the real split for the paper's factor under “The result”.

  1. Why do some stocks earn more than others?The standard answer is the Fama–French . It explains most of the differences in returns with five factors. Each is the return of one group of stocks minus another.
  2. One of the five factors measures profitability.RMW (robust minus weak) is the return of the most profitable firms minus the least profitable. To build it, every firm is ranked against the whole market, separately for large and for small firms.
  3. Whatever the five factors leave unexplained is called alpha.A strategy that keeps earning has found something the model is missing. Alpha is the measure I use to judge the result later on.
03 / The limitation

One ranking for the whole market.

Fig 4
GrocersChip makersSoftware firms

Twelve hypothetical firms. Figures are profit margins in percent, chosen for illustration.

  1. Take twelve firms in three industries.Margins vary much more from one industry to the next than they do inside an industry.
  2. A market-wide ranking mostly sorts industries.The top third is entirely software and the bottom third is entirely grocers. The profitability factor buys the first group and sells the second.
  3. Individual firms end up in the wrong group.The best-run grocer lands in the group that gets sold, and the weakest software firm lands in the group that gets bought.
  4. Ranked against its own peers, each firm lands where it belongs.The top third now holds the leader of each industry, and the two firms trade places.
04 / A real case

VMware, June 2015.

Fig 5

Redrawn from Figure 1 of the paper, which names five of the 25 peers.

  1. Against the whole market, VMware looks unremarkable.Its of +0.048 puts it in the middle 40% of a market-wide sort, so the factor leaves it out of both the long and the short side.
  2. Against its peers, VMware stands out.Its 25 peers are loss-making enterprise-software firms with a median of −0.070. VMware sits +0.118 above them, which puts it in the top 30%. Both views use the same firm in the same month. The comparison group is the one thing that changed.
05 / The idea

One line changes the sort.

Fig 6
level = peer median + deviationTrue for every firm, by construction
Peer-common

The business the firm is in

The profitability a firm shares with others facing the same demand, costs and regulation. RMW prices mostly this part.

Firm-specific

How the firm runs that business

What remains once the peer median is removed. The five-factor model has no factor for it.

Fig 7

  1. Any firm's profitability splits into two parts.The first part is the . The second is how far the firm sits from that median. The two always add up to the firm's own number.
  2. The five-factor model prices mostly the first part.Across the market, most of the variation in profitability comes from differences between industries, and that is what RMW picks up. I test whether investors are also paid for the second part, the firm-specific deviation. Everything else in the Fama–French construction stays as it is.
  3. Try the transform yourself.The sort ranks firms on the deviation alone. The VMware preset uses the values from the paper, and the other two presets are illustrations.
06 / The peers

Businesses change faster than their industry codes.

Fig 8
Company
Year

  1. Industry codes are assigned once and rarely revisited.74% of firms in the sample never change their four-digit . Amazon's has read 5961, catalog and mail-order retail, since its first annual report. In 1999 its text-identified peers were mostly internet and software firms.
  2. Peers identified from filing text follow the business.By 2010, 13 of Amazon's 25 closest peers were retailers.
  3. They keep following it as the business grows.By 2024 the peer set spans retail, software and hardware, while the SIC code still reads 5961.
  4. Does this only happen at large firms?I looked at firms that were never large. REX American Resources ran electronics stores and now produces ethanol, and its peers moved from retailers to fuel and chemical producers. Its market value is about $800 million.
  5. The filing text can pick up a change years before the code does.8x8 went from video chips to internet telephony to cloud software. By 2005 only one of its 25 peers was still a hardware maker, yet its SIC code said semiconductors until 2007.
  6. Some codes never catch up.Spok ran paging networks and now sells hospital communication software. Its SIC code still reads radiotelephone communications.
  7. A stable business keeps a stable peer group.Arrow Financial is a small bank, and all 25 of its peers are banks in every year shown.
07 / Firm size

Is this a large-firm effect?

Fig 9 New for this page
Smallest firmsSmall firmsLarge firms

This analysis is new and does not appear in the paper. I computed it from the paper's monthly peer files, using June snapshots from 1996 to 2025 and NYSE size breakpoints. The appendix describes the method.

  1. I first measured how much peer groups move in a single month.About 2% of a firm's peer mix sits in a different industry one month later, using the same model. I treat that as the baseline.
  2. Over several years the movement is much larger.After five years about 27% of the peer mix is in different industries. Some of that comes from retraining the model each year, and this measure cannot tell that apart from real change in the business.
  3. Small firms move as much as large ones.After ten years the figure is 31% for the smallest firms, 30% for small firms and 32% for large firms. The 0.4-point gap between the smallest and the largest is well inside the margin of error.
08 / No look-ahead

Peers without hindsight.

The corpus
1.03M

Filings read: 297,827 annual 10-Ks and 731,470 quarterly 10-Qs

576

Dimensions in each firm's embedding

25

Peers per firm, by cosine similarity, refreshed monthly

37.6M

Firm, peer, month observations across thirty snapshots

Fig 10

Fig 11

Share of a focal firm's 25 peers in each SIC division. Rows sum to 100%.

  1. Peers come from the text of regulatory filings.A language model reads what each firm writes about its business, risks and competition in its filings and turns it into a point in space. A firm's peers are its 25 nearest neighbours, recomputed every month.
  2. Why not use an off-the-shelf language model?Models such as GPT or BERT were trained on text through 2024 or later. Ask one for a firm's peers in 2003 and it already knows which firms merged, failed or changed course. Researchers call this , and it can make a backtest look far better than it should.
  3. So I trained my own models, one for each year.There are thirty of them, covering 1996 to 2025. Each has about 57 million parameters and the same design, and each is trained from scratch on filings accepted before its year.
  4. No model has seen a filing from its own year or later.The 2025 model has read 1993 through 2024. The training text grows by one year at a time, which builds the cutoff into the training itself.
  5. Do the resulting peers make economic sense?Manufacturing and finance firms keep 86% and 87% of their peers inside their own division. Overall agreement with two-digit SIC codes is 0.59. The peers match the code where an industry is coherent and move away from it where a firm's real competitors sit in another industry.
09 / The test

Three decades, one held out.

Fig 12

Hover or tap a marker for the event.

Fig 13

  1. I made every design choice on the first decade.The sample covers 354 months, from July 1996 to December 2025. The training window runs through 2005.
  2. I used the second decade to validate those choices.It runs from 2006 to 2015 and includes the financial crisis.
  3. I set the third decade aside.The years 2016 to 2025 were held out before the headline factor was built, which makes them an test. Each of the three windows contains a full market cycle and a major drawdown, so the test doubles as a stress test.
  4. How much can a single decade show?Less than I would like. A factor with true S has an expected of about S × √years, so ten years is often too short to reach significance. I draw my conclusions from the long windows and use each decade to check that the sign and the size hold up.
10 / The result

A five-factor alpha of 4.57% a year.

Whole sample, 1996 to 2025
PR_OPq, quarterly+4.57%

Five-factor alpha per year. t = 5.00. Sharpe ratio 0.78.

PR_RMW, annual+3.51%

Five-factor alpha per year. t = 3.39. Sharpe ratio 0.73.

Fig 14
Explained by the five factorsAlpha

Percent per year. The 4.7% is the paper's mean monthly return of 0.39%, annualized.

Fig 15
Benchmark
PR_OPqPR_RMW
  1. The factor buys firms that out-earn their peers and sells those that trail.This is a , and I build it two ways. PR_RMW is rebalanced once a year. PR_OPq is refreshed after each quarterly earnings report.
  2. Almost all of the return is alpha.PR_OPq earned about 4.7% a year, and the five factors account for only about 0.1 of that. For comparison, RMW earned 3.9% over the same period, and the model explains all of it because RMW is one of the five factors.
  3. The alpha shows up in both decades I used to design the factor.It is +6.08% a year in the training decade (t = 3.84) and +3.84% in the validation decade (t = 2.60).
  4. It also shows up in the decade I held out.From 2016 to 2025 the alpha is +3.38% a year (t = 2.42). Over the whole sample it is +4.57% (t = 5.00), which clears the t > 3 threshold proposed for new factors.
11 / Relation to RMW

A different half of profitability.

Fig 16
Raw quarterly profitabilityPeer-relative
Fig 17
Fig 18
Evaluated on

  1. Is this just RMW under another name?I checked how closely each version follows RMW. Raw quarterly profitability has a of +0.84 on RMW. The peer-relative version has a loading of +0.26, which is 69% lower, and its alpha is higher. The correlation between the two factors is 0.66.
  2. Its returns are about half as volatile.RMW has a monthly volatility of 2.80% and a worst month of −19%. The peer-relative factors come in at 1.39% and 1.72%, and the Sharpe ratio rises from 0.41 to 0.73 and 0.78.
  3. Does adding it improve the five-factor model out of sample?I estimate the best mix of factors on the training decade and apply it, unchanged, to the held-out decade. With the five factors alone, that mix earns a Sharpe ratio of 0.63.
  4. Adding the peer-relative factor raises the Sharpe ratio to 0.88.With both peer-relative factors it reaches 0.91. A copy of RMW would add nothing here, because the model already holds RMW. The gain shows that the factor carries information the five factors lack.
12 / Stability

Positive in almost every five-year window.

Fig 19
PR_OPq, quarterlyPR_RMW, annual
  1. Rolling five-year alpha is positive in 294 of 295 windows.That count is for PR_OPq. PR_RMW is positive in 282 of 295. The alpha is highest in the earliest windows, and a structural-break scan finds a single break in March 2001, at the end of the dot-com period. The results are stable after that date.
13 / Where it helps

The gain follows what the model already prices.

Fig 20
Raw loading on

Fig 21
Order by
RawPeer-relative
  1. Does the peer comparison help other characteristics too?I applied the same transform to 21 accounting characteristics. Each point is one of them, measured on the training decade. The horizontal axis is its raw loading on a factor, and the vertical axis is the alpha it gains from the peer comparison.
  2. The gain is larger where RMW and HML already load on the characteristic.The slope is +9.28 on the RMW loading (t = 5.35, R² = 0.60) and +10.00 on the HML loading. The size and investment loadings have no predictive power. With only 21 points, I read this as a pattern in the data and stop short of calling it a mechanism.
  3. About half of the characteristics gain.Over the whole sample, R&D intensity and SG&A intensity gain 7.7 and 5.8 points of annual alpha. Cash holdings lose ground. The chart reports all 21.
14 / The limits

Where the idea stops working.

Fig 22
WorksAlpha not different from zero
Fig 23
Median

  1. The factor works when the data is fresh.The quarterly version earns +4.57% a year (t = 5.00). The annual version, rebalanced once a year, earns +3.51% (t = 3.39).
  2. The same recipe fails on value and on stale data.Peer-relative book-to-market earns −0.53% (t = −0.71), and annual profitability rebalanced every month earns +0.19% (t = 0.19). One likely reason for the first result is that a firm's share price moves closely with its peers' prices, so little is left once the peer median is removed. I have only these two cases, so I treat the explanation as tentative.
  3. Does the alpha come from holding smaller firms?The long leg does tilt toward smaller firms, so I tested that directly. A peer-relative size factor shows an alpha of +6.1% (t = 2.6) in the training decade, and it disappears out of sample. I leave open whether the profitability premium reflects risk or mispricing.
Conclusion

The strength of a firm's profitability is a comparison.

When firms are ranked against the whole market, profitability mostly shows which industry each one is in. Once they are ranked against peers identified from filing text, with no look-ahead, the ranking picks up how well each firm is run. The five-factor model has no factor for that part, and adding one improves it. The same peers can now be applied to every other characteristic investors use to rank firms.

Cite the paper

Celik, Neil. 2026. “Profitability, Relative to Peers: Text-Identified Reference Groups and the Cross-Section of Returns.” Working paper, The University of Chicago Booth School of Business, June.

Appendix

Tables and method notes.

The numbers behind the figures, and a description of how I ran the firm-size analysis.

Five-factor alpha by window (Fig 15)
Maximum Sharpe ratio by model (Fig 18)
What predicts the gain from peer differencing (Fig 20)
All 21 characteristics (Fig 21)
Factor variants (Fig 22)
Peer-group change by firm size (Fig 9), and how it was measured

This analysis is new and does not appear in the paper. It uses the paper's own monthly peer files, which list each firm's 25 peers every June from 1996 to 2025.

The measure is the total-variation distance between the two-digit SIC mix of a firm's peers at two dates. I hold each peer's industry label fixed, so the measure moves when the set of peers changes. Size groups use New York Stock Exchange breakpoints: smallest is below the 20th percentile of market value, small is the 20th to the 50th, large is above the median.

A new model is trained each year, so some of the change comes from the model and some from the business, and this measure cannot separate the two. The comparison across sizes is unaffected, because every size group is read by the same models. At ten years the smallest-minus-largest difference is −0.4 points with a firm-clustered standard error of 0.7.

The four small firms shown in Fig 8 were chosen as clear illustrations from firms that stayed below the NYSE median size in every year observed. Among 1,326 such firms with at least 15 years of history, 17% had more than half of their peer mix replaced between their first and last three years.

Cover