The Peer Group: Choosing Comparables Honestly

A peer group is a published claim that these companies are comparable for one stated question. Choosing it honestly means fixing the set before the answer is visible, writing down who was left out and why, and stating the reading across every defensible set rather than the one that reads best. The set, and not the companies in it, usually decides how a comparison reads.

The opening sequence of this subject area covers how a comparison set gets built: the six steps, the four tests a candidate has to pass, and a warning about circularity. One step later comes the moment a set stops being scaffolding on an analyst's desk and becomes a sentence somebody else reads. Crossing that line changes the set's standing, its claim and who is answerable for it, and almost nothing written about comparables ever mentions that the crossing happens.

What changes when a comparison set goes into a document?

On the desk, a comparison set is a working choice. The only person relying on it is the analyst, who can remember why, so a company can be added on Tuesday and dropped on Thursday and nothing at all has happened. The moment the same set appears in something published, three things change at once. The set becomes a claim. The set becomes a claim made to somebody who cannot see how it was assembled. And the set becomes a claim the analyst will be asked about later, by a reader, by a reviewer, or by the analyst in eight months when the call has gone the wrong way.

The everyday shape of it is familiar. A school tells a parent that their child came third. Third out of what, and who decided the out-of-what? If the comparison is against the whole year group, third means one thing. If it is against the six children who sat the optional harder paper, it means something quite different, and the number three is identical in both. The number does not carry its own set, so a reader who is handed the number and not the set has been given something they can believe or disbelieve but cannot check.

Most published comparisons put a reader in exactly that condition. A sell-sideWhere research gets published for other people to trade on, by a house paid for producing the research rather than for running a book of its own. note shows a table of multiplesA price written as so many times some figure from the accounts, used to line two companies up quickly. How one is built is taught elsewhere in this subject area. and a line saying the subject trades below its peers. The peers are named, sometimes. Whoever was considered and left out is named essentially never. So the reader inherits the whole construction and can see about a third of it.

Where does honesty actually live in a comparison?

Ask most people what an honest peer group looks like and they will describe a membership: the right companies, the ones that really are similar. The membership answer cannot survive five minutes of pressure. Honesty sits somewhere else entirely, and seeing why it does is worth the five minutes.

Take two careful analysts, both experienced, both acting in good faith, both asked the same question about the same company. One argues that only a maker with a comparable customer base belongs in the set. The other argues that only a maker facing a comparable input basket belongs. Both arguments are sound. The two arguments produce different sets. Both people are honest. Since two honest people can arrive at two different memberships, honesty cannot be a property of the membership, and any test that looks only at who is in the set is testing the wrong object.

So where does it live? In the record of how the set came to be. Three things, and they are all about timing and disclosure rather than about judgement:

The three tests have something in common. None of them asks whether the choice was a good one. All of them ask whether somebody else can see how the choice was made and when. The reason is that judgement is not auditable and process is, and a discipline that can only be enforced by trusting the judgement of the person doing it is not a discipline.

How does a set go wrong when nobody lies?

A set goes wrong most often through a mechanism in which no single moment feels like cheating. Almost nobody defends against it for that reason.

Nobody sits down and picks a dishonest set. Iteration happens instead. An analyst builds a set and computes the comparison. The answer looks odd, or awkward, or simply unlike what was expected. So the membership gets looked at again. The membership is looked at properly, with real care, and something genuine turns up: one company sells to a completely different customer, or reports on a basis that does not line up. That company comes out, for that real reason, and the comparison is recomputed. Every single step in that sequence is defensible on its own, and a reviewer who examines any one of them will find nothing wrong.

The sequence is not defensible, and the reason is a matter of timing rather than of reasoning. The only thing that triggered the reconsideration was the answer, and the difference eventually found was equally true before the comparison was run and triggered nothing then. The finished document records the reason and not the trigger, so the whole of the problem is invisible in it.

The everyday version is familiar to anybody who has weighed themselves. The scale reads badly, so it must be uneven, and it gets moved to the tiled part of the floor. A second weighing follows. Was the tiled floor a better surface? Probably. Would the scale have been moved if the first reading had been the welcome one? The scale would have stayed where it was. The reason is genuine and the process has still stopped measuring weight and started measuring preference.

The test that catches this is short enough to hold in the head. Would that reason for the change have been given before the answer was seen? If yes, the change is made and the date noted. If no, or if the answer cannot honestly be told, the change is being driven by the result and both readings have to be published, the one before the change and the one after, with the change and its trigger recorded between them.

What does an exclusion record have to carry?

An exclusion record is the written list of every candidate considered and not used, with a reason against each name. An exclusion record is the single most useful document in this whole area and is almost never kept. Keeping one looks like paperwork about companies nobody is writing about.

To be worth anything it has to contain the following. Every candidate that was genuinely considered, not just the near misses that stayed in memory. The reason each one was left out. And, critically, a reason category, taken from a short fixed list rather than written freshly each time. Categories that carry real weight are: a different customer, a different cost driverThe one or two input costs that decide most of what a product costs to make, so a swing in them shows up quickly in the profit ladder., a different stage of the chain, a different capital intensityHow much plant and machinery a business has to fund to earn a rupee of sales, which is why two makers of similar size can need very different amounts of money behind them., and disclosure the filings never carry.

Not comparable is not a reason, it is the most common entry in every exclusion record that has ever been kept, and requiring a category from a fixed list is precisely what converts the record from a formality into a check. The mechanism is simple: a free-text box accepts anything, and a fixed list forces a statement of which kind of difference is being relied on. The moment a category has to be picked, half the exclusions that were about to be waved through stop fitting any of them, and that is the record earning its keep.

The case record supplies an instance. Sarvani Coatings Limited makes decorative paints and industrial coatings. Thottam Chemicals Limited supplies resins and additives. Is Thottam Chemicals a comparable? No, and the reason has a category: a different stage of the chain. Thottam Chemicals sells into makers like Sarvani Coatings rather than alongside them, so its input costs are its suppliers and its selling prices are Sarvani Coatings costs. Written that way the exclusion is checkable by a reader in one line. Written as not comparable it is checkable by nobody.

Can two fixed companies really give four different answers?

Now the demonstration, and it is built so that nothing about the companies moves. One question, two comparison companies, both of them in every construction, and not a single figure altered anywhere.

The question: is Sarvani Coatings Limited's gross marginThe slice of every sales rupee that survives the bill for physical inputs, and only that bill. Wages, freight and the rest come out further down. gain of 3.0 percentage pointsThe unit for the gap between two percentages. Moving from 43 per cent to 46 per cent is three percentage points, not three per cent. over the two years from year one to year three shared with the field, or particular to the company? Every gain in this comparison is that same two-year move, year one to year three, and a one-year move is a different measurement that cannot be set beside it.

The two comparison companies, both from the case record. Nandivarman Paints Limited, larger at an illustrative Rs 14,490 crore of revenue and almost entirely decorative, whose gross margin rose 2.4 points over the same two years. Kesaria Surface Solutions Limited, smaller at an illustrative Rs 1,449 crore and industrial heavy, whose gross margin rose 3.6 points over the same two years. Both are defensible members and both stay in for all four constructions below.

People trust the revenue weighting most and inspect it least, so the revenue weighting is worth writing out. The two revenues used are Rs 14,490 crore and Rs 1,449 crore, summing to Rs 15,939 crore. The case record says only that one comparison company is larger and one is smaller, so both revenues are illustrative sizes rather than record figures. The weights the two revenues produce are 90.91 per cent and 9.09 per cent. The field gain is 14,490 times 2.4 plus 1,449 times 3.6, all over 15,939, giving 2.5091 points and printing as 2.51.

Four ordinary constructions of two fixed companies place the same subject ahead, behind, level and ahead again, and the entire movement came from decisions about the set rather than from anything any company did.

How much of the answer did the set decide?

Hold both companies in the set permanently, so membership is no longer in play at all, and change only the basis of the average. Give the larger member a weight of nothing and the field gain is 3.60 points. Give it the whole weight and the field gain is 2.40 points. A weighted average of two fixed numbers is a straight line in the weight, so everything in between falls on that line.

The weighting alone produces the central finding. With membership fixed and both companies permanently in the set, the basis of the average alone spans the entire 1.20 point range that the membership choice spanned, so the two decisions people treat so differently have exactly the same power over the answer. Adding and removing companies feels like a serious editorial act. Choosing between equal and revenue weighting feels like a housekeeping detail. The two decisions move the number by the same amount.

The two ordinary weightings sit at 50.00 per cent and 90.91 per cent on the larger member and give 3.00 and 2.51 points, a move of 0.49 of a point from one perfectly standard choice to another perfectly standard choice. The weight at which the field genuinely equals the subject, worked against the recomputed 2.9607 rather than the rounded 3.0, is 53.28 per cent. A weight of 53.28 per cent is not equal weighting and is not any weighting anybody would have picked on purpose.

What has to be published with a comparison?

Five items, and a comparison missing any of them cannot be reproduced by the person reading it.

Equal weighting and revenue weighting are both entirely ordinary choices and they do not agree with each other, so the basis of the average is the item everybody omits and the one that moves the answer most quietly. Nobody hides the basis. Picking a basis simply never occurs to anyone as a decision at all, and never occurring to anyone keeps something out of view far more effectively than concealing it would.

What gets published when defensible sets disagree?

The range is published, every construction that produced it is named, and what would narrow it is stated. Publishing the range beats the alternative, and being clear about why it beats the alternative matters more than sounding cautious.

The alternative is to pick one construction and print its number. Whichever one is picked, the reader receives a figure that looks like a measurement and is actually a measurement plus a choice, with the choice removed. The reader cannot tell that a choice was made, so cannot disagree with it, so cannot use their own judgement anywhere. A reader who receives the range can act on it, and a reader who receives the single figure can only accept or reject it. A range across defensible constructions is therefore both a more honest output than a single figure from the construction that read best and a more useful one.

For this comparison the honest output reads: over the two years from year one to year three, the field gained between 2.40 and 3.60 percentage points of gross margin depending on the construction, the four constructions are named with the basis of each, and Sarvani Coatings gained 3.0 points as published, or 2.9607 points worked from the rupee absolutes. Two things would narrow the range: a fourth comparable maker of similar scale, or segment disclosure from either comparison company that let a decorative-only comparison be run on both. Naming what would narrow the range is the part that turns a hedge into a plan.

What did every construction still agree about?

Four constructions that disagree can look like an exercise worth nothing. The exercise was not worthless, and the reason matters more than the disagreement does.

Every one of the four constructions agrees that all three companies gained gross margin over the two years. Sarvani Coatings gained 2.9607 points worked from the rupee absolutes, Nandivarman Paints 2.4 points, Kesaria Surface Solutions 3.6 points. Not one construction produces a field that went backwards, and not one produces a subject that went backwards. The disagreement between constructions sits on top of an agreement, and the agreement is real evidence that narrows the question even though it cannot rank the companies.

The shared direction rules something out. If margins across this field had been flat and only Sarvani Coatings had gained, the explanation would have to be particular to the company. Margins were not flat. Every maker in the set gained. A pricing environment that let makers price ahead of their input costs fits that pattern, and so does an account in which Sarvani Coatings simply priced better than the others, and the comparison cannot separate the two. Removing one explanation and leaving two standing is what real evidence usually does.

One arrangement in the numbers looks like a result and is not. Sarvani Coatings 3.0 points sits neatly between the peer gains of 2.4 and 3.6, and the equally weighted average of the two lands on exactly 3.00. The neatness is a coincidence. Nothing about a market makes a subject sit midway between its comparison companies, and an average landing on a round figure is arithmetic rather than evidence.

What if there is no set to build at all?

Sometimes the honest answer is that the comparison cannot be made. If the only genuinely comparable makers are unlisted, or listed but disclosing nothing at the level the question needs, then no defensible set is available, and building one anyway manufactures a comparison out of companies already judged not comparable.

Where no defensible set exists, the honest output is the company's own record over time, and saying that the set could not be built is a finding rather than a failure. A reader told so is better informed than a reader handed a comparison against three loosely related makers. The first reader knows what they do not know and the second does not. Nothing in a spreadsheet issues a warning: a comparison against a forced set computes exactly as cleanly as a comparison against a real one, and prints to the same number of decimal places.