Twins
Remember valuation class? Yeah, I do. "Here's a valuation tool you plug and play for every stock you will ever pitch your college investment group and indefinitely into your career." Boring, in my opinion. Why are we valuing equities like bonds through the assumption they have regular free cash flows extending 5Y into the future? Why are comparable companies "comparable" to an equity you have a particular interest in? Why are we calling the gap between the median multiple of "comparable" companies and your company "upside"? It's widely used on the Street, and the whole thing hinges on one step, which is picking the comps. This usually means grabbing whoever shares the stock's GICS code and trimming the list until it feels right.
The problem I have with that is that an industry code describes what a company sells, not what kind of a business it is financially. A capital-light and high-margin business has more in common with companies with those same characteristics than a struggling competitor in the same GICS bucket in my opinion, at least in terms of how the market prices it. So, I'm extrapolating on an idea back from school - how do we build a model that scores companies on shared characteristics, measures how far apart each pair of companies sits, and then ranks every company's closest twins based on the financials?
I'm doing this for the S&P 500 to see how this idea works across industries and across time. The original was an Excel model that did this for one company and one date, but nonetheless turned the cluster of twins for a given company into a price target.
A price target doesn't prove much, especially from one date, so this model got rebuilt as a backtest of the concept. Using FactSet data at 28 dates from January 2020 to April 2026, I found every company's twins at each date and tested two main hypotheses:
H1: Do twins actually move together over the next 12 weeks and do they beat industry peers at the same thing?
H2: If a company is cheap relative to its twins, does it outperform over the next 12 weeks?
Here's some stats to get the ball rolling:
The Theory
Relative valuation is the law of one price, basically. If two companies have the same growth, the same profitability, and the same risk, they should trade at similar multiples. When one doesn't, the market either knows something or the gap should close (that's where more fundamental analysis comes in to see if the market really does know something). The problem is that the goal is to compare companies of similar growth and risk, but industry codes/arbitrary comps lists are somewhat of a cheap way out of doing the work. My results show that how you pick the peers matters more than the multiple you end up using.
Bhojraj and Lee (2002) showed that choosing peers by their "warranted multiple" (what their fundamentals say they should trade at) beats choosing them by industry. Further, Bhojraj, Lee, and Oler (2003) found that among industry schemes, GICS is the best at grouping companies that move and get valued together, which is why it ended up as my benchmark for this study and not some weaker classification. More recently, Hoberg and Phillips (2010) built industries out of product descriptions in 10-Ks, and Lee, Ma, and Wang (2015) built peer groups from companies investors look up together on EDGAR, and both styles beat GICS. On the multiples side, Liu, Nissim, and Thomas (2002) found that forward earnings multiples do the best job of pricing a company off its peers, which is why every multiple here uses the NTM consensus estimates rather than trailing numbers.
The fingerprint is my version of picking peers by fundamentals. Instead of assuming that industry codes do the work for you, it lets the peers fall out of the math (no matter where they came from).
What Goes Into "a Twin"
The 12 metrics that I chose aren't random, they are each a characteristic from the asset pricing literature. More importantly, the literature ties these metrics to how a stock is valued and how it returns. They are grouped into 5 main families:
Profitability: gross profit over total assets (Novy-Marx 2013's "other side of value"), operating margin, FCF margin, and ROE.
Investment: total asset growth over one and three years, sign-flipped so conservative companies score higher (Cooper, Gulen, and Schill 2008 show that fast asset growers underperform).
Earnings quality: total accruals, or net income minus operating cash flow over assets (Sloan 1996).
News: earnings and revenue surprise against consensus, since post-earnings announcement drift is one of the oldest anomalies around (Bernard and Thomas 1989).
Value and leverage: book-to-market, debt-to-equity, and debt-to-assets.
Profitability and investment are the two factors that Fama and French (2015) added to their model, so a lot of the fingerprint lines up with the five-factor world intentionally!
For each date, every metric is turned into a z-score across the S&P 500, and the distance between two companies is the Euclidean distance between their 12 z-scores:
where the sum runs over the nij metrics both companies actually report, and the 12/nij term scales it back up. This is so a pair missing data doesn't look artificially close. A pair needs at least 8 shared metrics to be compared at all. Similarity is just 1/(1 + d), so identical fingerprints score 1 and it decays to 0 from there. A company's "cluster" is itself plus its k nearest twins.
Before backtesting, the original Excel model needed some fixing. The above chart is the reason why, since the orange bars are the classic z-score (used by the first version). The problem is that one crazy value blows up the standard deviation and crushes everyone else together. Three-year asset growth had a company at −186,295 (a tiny asset base three years earlier), earnings surprise had readings past ±3,000%, and ROE had a company at 31,560% (lol). The result is that on earnings surprise the middle half of the S&P 500 sat within 0.06 of each other, and on ROE within 0.07, which represents how 4 or 5 of the dimensions were pretty much useless.
The indigo bars are the better version: subtract the median, divide by the median absolute deviation (scaled by 1.4826 so it's comparable to a standard deviation), and cap the result at ±3. Now every metric spreads companies about equally (IQR of 1.35-1.58), so all twelve actually count for something. The other (smaller) fixes were just missing-data scaling in the distance formula. Before that, companies like GEHC and KVUE that didn't exist on some of the earlier dates of the backtest showed up as XOM's two closest twins in 2020. Having one data point in common made them look identical, even if they were non-existent. Finally, a 90-day lag was placed on all annual data so a January fingerprint can't use a 10-K that wasn't filed until February.
This chart is VERY important in changing the way you think about comps. Take XOM for example - its eight closest fingerprints are EPAM, Cognizant, Cooper, Block, L3Harris, IDEX, Ingersoll Rand, and Textron. Not a single oil company. Apple's are just as different with the lineup including Gilead, Union Pacific, Netflix, Moody's, IDEXX, KLA, eBay, and Autodesk. These companies don't have much in common as businesses but they do financially (high margins, high ROE, low book value from years of buybacks, and steady asset growth). JPMorgan's twins are mostly financial-ish (KKR, BNY, AmEx, Northern Trust, State Street, NextEra, NXP, and General Mills), and this is because banks, custodians, and asset managers share a leverage profile. But, again, none of these are in the banking industry.
Costco is the exception that proves the rule: Walmart and Target (shown as orange since they're in the same GICS industry as Costco) come up first because big-box retail has a very specific fingerprint (thin margins, high asset turnover, and low accruals). So the fingerprint can find business twins when a business model leaves a specific "financial signature". Most of the time, the model finds style twins: companies with the same mix of profitability, investment, value and leverage. Whether that's useful is exactly what H1 tests.
This is the original graphic from the Excel model I made, rebuilt with fixed methodology. The idea behind the cluster premia is that you take the target's k nearest twins, take the median of their NTM multiples, apply that median to the target's own NTM estimate, and then bang - there is your price. For the enterprise-value multiples that means (multiple × EBITDA or EBIT − net debt) over market cap. For P/E it's the multiple times EPS over the share price. For P/FCF and P/S it's the multiple times the metric over the market cap. The blended panel, shown in orange, is an equal-weighted average of whichever of the five are available. Each dot is one cluster size, all the way from a two-company cluster (ex: XOM and its nearest twin) out to a cluster size of 100.
The shape is what matters here. With tiny clusters, the upside jumps around. This is because one twin's multiple is the whole answer (XOM's P/FCF-implied upside is +149% with two to four twins and 0% with five). By 10-20 twins the medians settle, and from there it barely moves. The blend lands at around an 8% upside at a cluster of 10 and at a 26% upside at 100. You can see that the multiples also disagree with each other quite a lot. XOM traded at 7.7x NTM EBITDA, which is well below its twins, so EV/EBITDA says +50% to +77%, while at 27.6x FCF, P/FCF says it's 28-42% expensive. That disagreement is normal for a capital-heavy business valued against capital-light twins, and it's probably a good reason to blend.
Well, of course the 18th of January 2020 was about three months before oil futures went negative, so this is a great illustration of why a single price target proves absolutely nothing and why the rest of the page devolves into backtesting instead.
H1: Do Twins Actually Move Together?
If the fingerprint captures what drives a stock's valuation, then a stock and its twins should respond to the same news. Maybe that's the same rate move, the same style rotation, the same earnings season. So, H1 has nothing to do with valuation at all. At each of the 28 dates selected, for each stock, I took the average return of its peer group over the next 12 weeks and checked how well it lines up with the stock's own return, across the whole cross-section. I compare four peer groups: fingerprint twins (the k nearest), GICS industry (every other member of the stock's industry, with there being ~67 industries represented in the S&P 500), GICS sector (11 sectors), and random (k random stocks, averaged over 50 draws).
Random is more of a sanity check, it should show zero. But if it doesn't, something might be wrong with the test. Industry is really what's important here. Beating random only shows that the fingerprint isn't noise, while beating GICS would prove that it's better than the thing everybody already uses. But it's also important to consider this: even if twins lose to industry, do they add anything to the industry bucket as extra? For that question I used a Fama-MacBeth "horse race" by regressing each stock's return on its industry average and its twin average together, date by date, and testing whether the twin coefficient stays positive.
Each point is one 12-week interval. For each stock, I compare its return to its peer group's average return. The point is the correlation across the ~490 stocks that qualify. The random line sits on zero the whole time, so the test shows that my hours making this weren't wasted. The twins (orange) average 0.24 and beat random on all 28 dates. So yes, twins are real. Companies with similar fingerprints move together as much as the companies in the same sector do on a rank basis (0.23 vs 0.21).
But the GICS industry (blue) wins on 24 of the 28 dates, averaging 0.40 against 0.24 for the twins (t = −6.6). The dark columns are the Covid crash window and the 2021-2022 rate shock. The pattern is the same in and out of stress, it seems. The four dates where the twins do win are June 2021, November 2021 (0.57 - the best twin reading in the sample, right as growth started suffering), December 2023, and October 2025. These are all stretches where style rotations dominated the tape. That's what you would expect if the twins are style twins, since they win exactly when the market is trading styles instead of industries.
Here's the horse race, and where there may be some merit behind this model. When each stock's return is regressed on its industry average and its twin average together, both coefficients stay positive on every single date. The industry slope averages 0.58 (t = 23) and the twin slope averages 0.41 (t = 12), with the twin slope never falling below 0.04. The share of the return spread explained goes from 17.2% with the industry alone to 20.0% with both.
Doesn't seem like this model is a worse industry classification, it's just a different one. Only 11% of a stock's ten nearest twins share its industry and 27% share its sector. So, the twins are picking up some co-movement that GICS doesn't really see. Value vs. growth, quality vs. junk, levered vs. conservative - this splits those differences for every industry. Seems like the best way to use this model is industry first, twins second.
Cluster size is the main knob from the original Excel model, so I wanted to see how every result depends on it in the new model. On the left you can see the twin correlation climbs pretty quickly from 0.16 with one twin to 0.23 at five, and then flattens around 0.25 from 15 twins and onward. Adding more twins averages away idiosyncratic noise clearly, but it doesn't get near the industry comparison's 0.40.
In the middle chart is the overlap with GICS. Even the single nearest twin is only in the same industry 18% of the time, and by 50 twins it's down to 6%. On the right is a bit of a sneak peek at H2, which is the predictive power of "cheap vs twins". It's pretty much flat at zero across every cluster size, whether the twins are chosen from anywhere or only within the stock's industry. Past ~20, the within-industry version becomes the whole industry, which is why that line goes flat. None of them beat plain value as seen by the dashed line, and plain value itself isn't really helpful/significant either.
H2: Does Being Cheap Relative to Your Twins Predict Returns?
This is the cluster premia chart turned into a trading signal. If comparables work the way the textbook says, then a company trading below its peers' multiples is mispriced and that gap should close. So, the cheapest names relative to their peers should beat the most expensive ones. There's a real argument for it, and I believe the notion that relative mispricing is the entire premise of pairs trading and most sell-side price targets. The Bhojraj and Lee work found that warranted-multiple peers do help forecast where a multiple is heading.
The argument against it is that a multiple isn't a price, but rather just a summary of growth, risk, quality, or whatever else. Anything the twins model doesn't capture shows up as a discount that's actually deserved (cheap for a reason). The 2020-2026 range was a rough stretch for value in general and 12 weeks may be too short for a valuation gap to close.
Given H1, the primary signal picks twins within the stock's own industry (k = 5), which gets you business similarity from GICS and financial similarity from the fingerprint. Three alternatives show up next: twins from any industry, the whole industry, and the whole market. The last one is just cheapness (a stock's multiple against the median of the S&P 500), so it's the baseline for this case that any peer-relativeness has to beat. Implied upside is computed exactly as it was in the cluster premia chart and blended equal-weight across the five multiples. Multiples on negative earnings are dropped and the financials sector is excluded since EV and FCF don't mean anything there. Each signal is scored three ways: (1) rank correlation with the next 12-week return (information coefficient), (2) the return of the cheapest fifth minus the most expensive fifth, and (3) a Fama-MacBeth regression against plain value.
If the signal worked, these bars would increase as you moved from left to right. Instead, there is a U-shape. The cheapest fifth relative to peers does the best (4.7% per 12 weeks for twins within the industry), but the most expensive fifth is right behind it at 4.1%, and the middle three sit between 2.7% and 3.0%. The same shape shows up for every peer group, including the whole market.
That doesn't look like valuation to me... The stocks that look the most mispriced against their peers in either direction are the ones with the unusual fundamentals, which means they are probably the most volatile, and the most volatile stocks do the best when the S&P 500 goes up (the average stock gained 3.4% per 12-week window during my observed time period). The spread between the ends of each chart is small and noisy, with +0.6% per quarter for the primary signal and +1.9% for value alone. Not really helpful/significant.
Here's the problem by period, and as you can see, the cheap-minus-expensive spread isn't a small and steady number. It's big swings that cancel out. In the Covid crash (as seen by the first dark column) cheap did very poorly (−14% for the primary signal and −18% for plain value) but then the value rotation paid it back with a +18% swing for the primary signal and +25% for plain value starting in December 2020 and November 2021 respectively. The second represents when the rate shock hit growth stocks. However, in 2023-2024, it comes back down as mega-cap growth started to take over again.
The important part is how closely orange tracks purple. The per-period spreads of the within-industry twin signal and plain market value are seen to correlate at 0.95. Comparing against twins takes out a piece of the value bet, which makes the swings smaller, but doesn't add any new sources of return. Compounded, $100 in the plain-value spread ends at $152, the twins-any-industry version at $131, industry peers at $115, and the primary within-industry signal at $111. These are gross, price-only, equal-weighted, and not something you could trade.
Last, every signal's average IC with a two-standard-error band. The primary blend's IC is +0.001 (t = 0.03), positive on 11 of 28 dates, and the single multiples split around zero. P/S is the best (or least bad) at +0.004, P/FCF the worst at −0.016. Measuring against industry-adjusted returns so the test only rewards picking winners within an industry (which is what peer-relative value claims to do) doesn't help much either (+0.007, t = 0.22). In the Fama-MacBeth regression with plain value alongside, the twin signal's coefficient is slightly negative (−0.02, t = −1.1). So, once you know how cheap a stock is against the whole market, knowing how cheap it is against its twins tells you nothing extra.
Well, put together, the two hypotheses make a nice story that is somewhat coherent. The twin model finds some structure since companies with similar characteristics do move together, and that co-movement is separate from each stock's respective industry and adds to it in each period. The market already prices that structure, however. Being cheap relative to your financial twins over a 12-week horizon is either deserved or a gap that doesn't close on any timeframe you could trade from your Robinhood account. For the original purpose of improving comps, I'd use the fingerprint to widen and check an industry peer set rather than replace it, and I'd treat a gap between twins as something to add to your equity research report rather than as an answer to open questions. The next steps are longer horizons where valuation gaps have more time to close, returns net of factor exposures, and peer groups based on text rather than numbers.
Method. The universe is the 500 current S&P 500 constituents at 28 formation dates spaced 12 weeks apart, from 18 January 2020 to 4 April 2026, with all data from FactSet. The twelve fingerprint metrics use annual (restated) fundamentals pulled as of 90 days before each formation date, so each date only sees reports that would have been public. Earnings and revenue surprises, NTM consensus estimates, prices, enterprise value and market cap are as of the formation date. Z-scores are robust (median and 1.4826 × MAD, capped at ±3) and calculated cross-sectionally at each date. Distances use shared metrics only, scaled by 12/n, with at least 8 shared metrics required. The forward return is the price change over the next 84 days. H1 uses k = 10 in the headline numbers (k = 5 and 20 give the same conclusions), random peers are averaged over 50 draws, and significance comes from paired t-tests on the 28 per-date statistics, which is valid without overlap corrections since the windows don't overlap. H2 uses k = 5 twins within the stock's GICS industry as the primary signal, with at least 3 valid peer multiples and 2 available multiples per stock. Each multiple's implied upside is winsorized at 1%/99% before the equal-weight blend, and Financials are excluded (about 370-410 stocks per date). The Fama-MacBeth regressions for H2 run on cross-sectional ranks. All the settings were fixed before looking at results.
Data considerations. The universe is today's S&P 500, so every company in it survived to 2026. That survivorship bias flatters anything long-cheap, since the cheap companies that went on to fail aren't in the data, which makes the H2 null more convincing rather than less. FactSet's annual data is restated, so even with the lag, some values reflect later corrections. GICS classifications are current, not as of each date, which matters around the 2023 reshuffle that moved Visa, Mastercard and PayPal into Financials. Forward returns are price only, which understates the return on high-dividend stocks by roughly a percent a quarter at most. With 28 non-overlapping periods, the tests have limited power, and a genuine but small effect (an IC around 0.02) would be hard to tell apart from zero. The H1 results are strong enough that this doesn't matter. For H2 it means "no detectable effect over 12 weeks," not "no effect ever." The single-company cluster premia chart (XOM, January 2020) is there to illustrate the method and isn't part of either test.
References. Bernard & Thomas (1989), Journal of Accounting Research; Bhojraj & Lee (2002), Journal of Accounting Research; Bhojraj, Lee & Oler (2003), Journal of Accounting Research; Cooper, Gulen & Schill (2008), Journal of Finance; Fama & French (2015), Journal of Financial Economics; Hoberg & Phillips (2010), Review of Financial Studies; Lee, Ma & Wang (2015), Journal of Financial Economics; Liu, Nissim & Thomas (2002), Journal of Accounting Research; Novy-Marx (2013), Journal of Financial Economics; Sloan (1996), The Accounting Review.