How to Select Comparable Companies: A Repeatable Six Step Screen

Comparable companies are selected with a written screen, not by feel. Six steps: put the subject in one testable sentence, list and count the universe the screen started from, apply the exclusions that need no judgement, fix four numeric bands, rank what survives, and record every name dropped beside the test it failed. The multiples are read only at step five. On six candidates this screen leaves three.

The whole procedure rests on one idea, and it is a procedural idea rather than a financial one: the screen has to be falsifiable by somebody other than the person who built it. Two things make a screen falsifiable, and only two: every test is a number or a yes-or-no question rather than a description, and every test is fixed before the multiples are visible. The first means a second person applying the same tests lands on the same set. The second means the tests cannot be quietly tuned until they produce an answer somebody already wanted. Everything in the six steps below is one of those two requirements wearing working clothes.

The same procedure runs whenever somebody works out what a flat should rent for. Nobody compares a two room flat above a shop with a four room house two streets away, and everybody knows why: size, locality and floor are what make two flats worth comparing. The move that costs people money is looking at the rents first and then deciding which flats count. Once the rents are visible, the rules a person picks are the rules that produce the answer already in their head, and they will not notice themselves doing it. The rules go down in writing first; the rents are read afterwards.

What are the six steps, and what order do they run in?

The six steps are define, list, exclude, band, rank, record. The six run in that order and the order is load bearing. Steps one to four are all completed before a single multiple is looked at, and step five is where the multiples first appear. That is the whole design. A screen where the multiples are visible from the start is not a screen at all: it is a set of names chosen for their answers, with criteria written afterwards to explain the choice.

Nothing in the six steps requires software. A sheet of paper carries a universe of six, four bands and a funnel perfectly well, and a screen assembled by hand is easier to check than one assembled by a query nobody can see. The steps do require that each one is written down as it is finished. A step completed in the head and never recorded is a step a second person cannot re-run.

Step one: what does it mean to define the subject so a test can read it?

Step one is one sentence about the subject, written so that every clause in it can be turned into a test. A subject definition that a screen can execute contains numbers and a plain description of the operations, and nothing else. Adjectives are not criteria. Well run, high quality and market leading cannot be applied to a candidate by a second person, so they cannot appear.

For Sankalp Industrial Systems Limited, invented, the sentence is this. A listed manufacturer of industrial valves and of precision castings, together with the aftermarketSelling spares, servicing and support for equipment that is already installed, rather than selling the equipment itself. parts and service that go with them, with revenue of Rs 12,00,00,00,000, an earnings before interest, tax, depreciation and amortisation (EBITDA) marginOperating profit before depreciation, taken as a share of revenue. Two companies with identical sales can sit ten points apart on it. of 24.0 per cent, forecast revenue growth of 10.0 per cent, a return on invested capitalWhat a rupee of capital locked inside the operation earns back each year, after tax, counting only what trading itself throws off. of 15.0 per cent, and net debt to EBITDABorrowings less cash, divided by one year of operating profit before depreciation. Roughly how many years of that profit it would take to clear the borrowing. of 1.67 times. Every one of those five is a number a test can be built on, and the business description is specific enough that a reader can say whether a candidate does that or something else.

Three words are missing from that sentence deliberately: good, cheap, growing. Each of the three is a conclusion about the subject rather than a description of it, and a screen opening on a conclusion goes out and finds the companies that agree with the conclusion. Step one describes; it does not judge.

Step two: what does stating the universe actually require?

Step two is a written list of every company considered, with a count at the top. A funnel with no starting number cannot be audited, so the count is written down before anything is removed. A reader handed a peer set of three has no way of knowing whether three came out of a starting list of six or a starting list of sixty, and those two screens are not remotely the same claim.

For this screen the universe is six companies in the same segment, all invented, and each takes a number here that the worked screen below uses. Number one is Aravalli Flow Controls Limited. Second comes Satpura Engineering Works Limited, largest of the candidates. Third comes Kaimur Industrial Limited, nearest the subject on size. Girnar Precision Limited takes fourth place. Shivalik Systems Limited is fifth, and its margin is the second highest here. Nallamala Components Limited closes the list as the smallest. Six is a small universe, and saying so is part of the record rather than an embarrassment to be hidden. A screen run over six names and a screen run over two hundred are both honest; only one of them can support a claim about a distribution.

Step two assumes nothing bought. A reader with no access to a paid data service can still build a universe by writing down the companies in the segment they can actually name, and then saying in the record how the universe was assembled. The failure is not a short list. The failure is a short list presented as though it were the whole market.

Step three: which tests are applied before any judgement at all?

Step three applies the tests that need no judgement, and it runs before the bands because a yes-or-no fact is cheaper to check than a number and because the two kinds of test mean different things when a name fails them. A hard exclusion is not a statement that a company is dissimilar; it is a statement that the number taken from it is the wrong kind of number. Five of them cover almost everything.

One, the company is not listedShares that can be bought and sold on an exchange, so a price is quoted every day rather than agreed once between two parties., so there is no quoted price to read. Sankalp Coatings Private Limited, the subsidiary of the subject itself, is exactly this case: a real business with real accounts and no observable price at all. Two, the accounts cannot be obtained, so the denominator cannot be built. Three, the reporting basis will not sit on the same footing as the subject's, so an operating profit figure taken from it is not measuring the same thing. Four, the reporting period cannot be lined up, so the year being compared is not the year intended. Five, the company is inside a live sale process, and a price agreed inside a sale carries a control premiumThe extra a buyer pays above the traded price to get the whole of a company and the right to run it. Buyers pay it; ordinary shareholders trading in the market do not. rather than a trading level.

Each of the five is answerable with a yes or a no by somebody who has never met the subject. Answerability by a stranger is the test for whether a criterion belongs in step three rather than step four. If answering it requires a decision about how similar is similar enough, it is a band, and it belongs later.

Step four: how are the four bands set, and when?

Step four writes four numeric ranges around the subject, one for size, one for profitability, one for growth and one for leverage, and it writes them before anybody looks at a multiple. A band is judgement made numeric: the decision about how similar counts as similar is taken once, in the open, and then applied to every candidate without further thought. A band is the only form of judgement a screen can carry, and it is the only form a second person can reproduce.

The rule that sets a band has to be stated in the same breath as the band itself, and one rule covers most cases. For a quantity that scales, use half to twice the subject. For a quantity that does not scale, use plus or minus a fixed distance. Revenue and growth scale, so they get half to twice. A margin is already a ratio and a leverage figure is already a ratio, so neither scales, and both get a fixed distance instead. The reason to state the rule and not merely the numbers is that the rule is what a second person argues with. A reader who disagrees with half to twice has something specific to disagree with; a reader handed only the edges has nothing to grip.

The timing is not a nicety. A band written after the multiples are on screen has been chosen with knowledge of what it will produce, and the finished output looks exactly the same either way. No arithmetic anywhere can distinguish a band set first from a band set afterwards. The only defence is the record.

What happens to a name that fails one band, and to one that fails two?

The exclusion rule is a count, not an impression. Two band failures and the name is out; exactly one and the name is borderline, decided in writing and reported both with it in and with it out. A company is never dropped because it felt wrong, and never kept because somebody liked it. The rule reads the number of failed tests and nothing else.

Skipping the borderline treatment is easy, and it is the step carrying the most weight. A name that clears three bands and misses the fourth sits exactly where the criteria stop being obvious, and no company in the universe is more informative. Dropping it silently hides that. Keeping it silently hides that too. Writing down that it failed one test, saying which, giving a reason for the decision, and then reporting the statistic both ways lets the next reader see the size of the choice rather than only its outcome.

Step five: why are the multiples read last?

Step five ranks what survived and only then reads the multiples. Reading the multiples last is not a courtesy to the procedure; it is the single thing that makes every earlier step mean something. Once a trading multipleA price the market is putting on a company right now, divided by one of that company's own figures, such as its sales or its operating profit. has been seen, no criterion written afterwards can be trusted, including by the person who wrote it.

Ranking is the smaller half of the step and it takes one line. Order the survivors by whichever dimension matters most, usually size, and the set arrives with a readable shape rather than as an unordered pile. Then the prices are read, on one date, and that date is written down. A multiple read on a Tuesday and a multiple read three weeks later are not members of the same set, and a set assembled across a fortnight of market movement carries a spread nobody chose.

Step six: what goes into the record?

Step six writes down six things, and the test of whether enough has been written is simple. Somebody who was not there, handed the record, should be able to re-run the screen to the same set without asking a single question. If they need to ask, the record is short by exactly the answer to that question.

The six lines are these: the universe and its count; the four bands as numbers, with the rule that set them; one line for each company naming the tests it passed and the tests it failed; the treatment applied to each borderline name and the reason; the date the prices were read; and the statistic computed on the clean survivors, on the survivors plus the borderline names, and on everybody. Not one line in that list is an opinion, and that is the point. A record made of decisions and counts can be argued with; a record made of adjectives can only be believed or disbelieved.

What does the screen look like run in full on six companies?

Everything so far is procedure. Here is the procedure carried out, once, end to end, on the six invented companies in the universe, and this is where the arithmetic lives. Run on these six candidates, the screen leaves three companies passing every band, two borderline names failing exactly one each, and one name excluded on three failures. Nothing in what follows is a judgement about any of them beyond the tests already written down.

Step one is the sentence already given for Sankalp Industrial Systems Limited. Step two is the universe of six, counted. Step three fires on nobody: all six are listed, all six publish accounts a reader can obtain, all six report on the same footing and over the same period, and none is inside a sale process. A step that catches nothing this time is still a step somebody has to be able to see was run, so the empty result is written into the record rather than skipped.

Step four sets the bands from the subject outwards, using the rule stated earlier. Revenue scales, so it gets half to twice. Growth gets half to twice, for the same reason. A ratio does not scale, and doubling a 24.0 per cent margin would produce a range nobody could defend. The margin gets plus or minus five points and the leverage figure gets plus or minus one turn.

Now apply those four numbers to the six candidates, one candidate at a time, without looking at anything else. Aravalli Flow Controls Limited has revenue of Rs 9,00,00,00,000, a margin of 20.0 per cent, growth of 4.0 per cent and leverage of 1.8 times: three inside, and growth outside the floor of 5.0 per cent. Satpura Engineering Works Limited at Rs 15,00,00,00,000, 21.0 per cent, 5.5 per cent and 2.2 times is inside on all four. Kaimur Industrial Limited at Rs 11,00,00,00,000, 23.0 per cent, 7.0 per cent and 1.5 times is inside on all four. Girnar Precision Limited at Rs 13,50,00,00,000, 24.5 per cent, 8.0 per cent and 1.2 times is inside on all four.

Shivalik Systems Limited is the interesting one. Revenue of Rs 8,00,00,00,000, a margin of 26.0 per cent and growth of 9.5 per cent all sit comfortably inside, and its leverage of 0.6 times misses the floor of 0.67 times by 0.07 of a turn. One failure, and the size of the miss changes nothing about the rule. Nallamala Components Limited fails three: revenue of Rs 5,00,00,00,000 is below the floor of Rs 6,00,00,00,000, a margin of 30.0 per cent is above the ceiling of 29.0 per cent, and it carries net cashCash on hand larger than borrowings, so the borrowing figure comes out below zero rather than above it. of 0.4 times its own operating profit before depreciation, nowhere near a floor of 0.67 times of net borrowing. Only its growth of 18.0 per cent lands inside a band.

The counts are what make the screen auditable, so write the funnel down as counts. Six in the universe. None of the hard exclusions fired, so six pass them. Three passing every band, two borderline on one failure each, and one excluded on three. A reader who is handed those four numbers can already tell that this screen was strict, that it was strict on a small universe, and that the strictness cost it half the starting list.

Only now are the multiples read, on one date, and written down. The three clean names, Satpura Engineering Works Limited, Kaimur Industrial Limited and Girnar Precision Limited, trade at 7.1, 7.6 and 8.0 times their operating profit before depreciation. The two borderline names, Aravalli Flow Controls Limited and Shivalik Systems Limited, trade at 6.6 and 8.7 times. The excluded name, Nallamala Components Limited, trades at 13.8 times. Sankalp Industrial Systems Limited itself trades at 7.78 times on the same measure, inside the spread of its own clean survivors and no further than that. Landing inside that spread settles nothing on its own. A reading in the middle of its comparables can mean the market sees the business those comparables describe, or it can mean two opposite misjudgements are cancelling inside one number. A multiple is a question rather than an answer, and it is worth exactly as much as the workings attached to it.

One detail is worth naming so that nobody reads a lesson into it. The excluded name here happens to carry the highest multiple in the set, at 13.8 times against a clean spread of 7.1 to 8.0 times. The coincidence is a property of this constructed universe and not a demonstration that the screen was right. A screen is defended by its record, never by the convenience of what it removed.

What is to be done when the screen leaves three names?

Three is uncomfortable, and the discomfort is the correct reaction rather than a problem to be engineered away. Four responses are available when a screen leaves three names, and three of them are acceptable. The three can be reported as a range, with the count stated. The borderline names can be brought in and the statistic reported both ways. One band can be widened, the widening recorded, and the figure reported both ways again. The three responses above are honest. The fourth, widening the bands until the count feels comfortable and reporting only the comfortable answer, is the one the whole procedure exists to stop.

Reporting all three answers at once is the response that costs nothing and buys the most. On the three clean names the readings run 7.1, 7.6 and 8.0 times, a range with a middle of 7.6 times. Carrying both borderline names in gives five readings of 6.6, 7.1, 7.6, 8.0 and 8.7 times, a median of 7.6 times and a mean of 7.60 times. The two coincide here. Including everybody, the excluded name included, gives a median of 7.8 times and a mean of 8.63 times. Three answers, one screen, all of them defensible with the record beside them.

How is a band widened without breaking the screen?

Widening is allowed. Widening quietly is not. Widen one band, write down which one, write down what it let back in, and report the statistic before and after. Change all four at once and nobody, including the analyst who made the change, can say which criterion produced the new set, and a set that cannot be attributed to a test cannot be reproduced from the record.

The answer is not what most people expect, so watch what single widenings actually do here. Take the revenue band from half to twice out to a third to three times. The widened band runs from Rs 4,00,00,00,000 to Rs 36,00,00,00,000. Nallamala Components Limited now clears the revenue test. And the set does not change at all: that company still fails the margin band and still fails the leverage band, so it still carries two failures and the rule still puts it out. A widening that admits nobody is a widening worth recording precisely because it shows how far the excluded name actually sits from the criteria.

Now take the growth band instead, from half to twice out to a third to three times. The widened band runs from 3.33 to 30.0 per cent. Aravalli Flow Controls Limited, whose 4.0 per cent growth was its only failure, becomes a clean name. The set goes from three clean names to four, the readings become 6.6, 7.1, 7.6 and 8.0 times, and the middle falls to 7.35 times. Widening the leverage band by half a turn instead does the mirror thing: Shivalik Systems Limited becomes clean, the fourth reading added is the highest of the survivors rather than the lowest, and the middle moves upward. Same universe, same subject, one band moved each time, and three different answers, each one attributable to exactly one decision.

How much does the width of one band actually decide?

The relationship between how wide a band is and how many companies survive it is the thing almost everybody underestimates, and it is easiest to feel by moving one edge and watching names appear. The number of survivors is a property of the band chosen, not a property of the companies. The prediction is committed to before anything moves.

Four settings are worth holding on to, and all four are in the drawing that follows. At plus or minus 10 per cent the band runs from Rs 10,80,00,00,000 to Rs 13,20,00,00,000 and one name passes, Kaimur Industrial Limited. At 25 per cent the band runs from Rs 9,00,00,00,000 to Rs 15,00,00,00,000 and four pass, with Aravalli Flow Controls Limited landing exactly on the lower edge and Satpura Engineering Works Limited exactly on the upper one. At 40 per cent five pass, and at 50 per cent five pass, that being the screen's own setting. At 60 per cent all six pass. Moving one number from 10 to 60 takes the survivor count from one to six without a single company changing in any way.

There is one more thing in that ladder and it is easy to miss. Two candidates sit exactly Rs 3,00,00,00,000 from the subject, one on each side of it: Aravalli Flow Controls Limited below at Rs 9,00,00,00,000, then Satpura Engineering Works Limited above at Rs 15,00,00,00,000. So the count jumps from two to four in a single step at 25 per cent. A symmetric band admits both in the same instant. On this universe the revenue test alone can never leave exactly three names, at any setting whatever. A screen that had to produce three would have to reach that number another way, and knowing so matters before anybody tries to tune a band to a count.

One honest wrinkle. The screen's own revenue band is not symmetric. Half to twice the subject runs from Rs 6,00,00,00,000 up to Rs 24,00,00,00,000. The control's plus or minus 50 per cent runs from Rs 6,00,00,00,000 up to Rs 18,00,00,00,000. The largest candidate in this universe is Satpura Engineering Works Limited at Rs 15,00,00,00,000, and no candidate sits in the stretch between the two upper edges. Both bands admit the same five names. The lower edge is doing all the work here. A size band usually behaves that way, and the habit is worth checking rather than assuming.

How does an analyst actually use this in a working week?

An equity analyst covering an industrial manufacturer does not build a peer set once. The analyst builds it, defends it in a review meeting, and then keeps defending it every time somebody senior names a company that is not in it. The record is what turns that conversation from an argument about taste into a check that takes ninety seconds. Somebody asks why a particular company is missing, and the answer is a line: it was in the universe and failed the margin band at 33.0 per cent against a ceiling of 29.0 per cent, or it was never in the universe because the subject sentence does not cover what it makes.

The same record does a second job further along. When a valuation is revisited three months later the prices have all moved, and the temptation is to rebuild the set from scratch. Rebuilding quietly produces a different set. Holding the bands fixed and re-reading only the prices leaves a record that can state exactly what moved. If a company entered or left the set between the two dates, the funnel shows it, and that movement is itself information about the segment rather than noise in the process.

A lender does something structurally identical when sizing a facility against comparable borrowers, and a person selling a small business does it when their adviser produces three comparable sales. In both cases the useful question is not whether the comparables look reasonable. The useful question is what the starting list was, what removed the rest of it, and whether the tests were written before or after somebody saw the answer. Anybody can ask those three questions, and they are the whole of what this procedure protects.