The Investment Thesis: Structure, Evidence, the Few Variables It Depends On, and How It Fails

An investment thesis is a single claim about a business, the two or three variables it depends on, the evidence that would break it, and the period over which it can be judged. A thesis without its disconfirming evidence written down first is a preference.

Three things arrive here already settled. How a quoted price is taken apart to read out the growth rate sitting inside it was worked through in the valuation material, and it is applied here rather than taught again. The published three year ladder for Sarvani Coatings Limited, an invented coatings manufacturer, and the unresolved question about where its gross margin gain actually came from, were carried across from the accounting and sector work. And the nine estimates that stood before the year was published, and the published outcome that replaced them, came from the earnings work. The thesis is the shape all of that gets poured into.

What is a Research Thesis, and when is a sentence not one?

Start away from markets. Suppose a neighbour says the vegetable seller at the corner will do better next year. The remark is a nice thing to say, and there is nothing in it to argue with. There is nothing in it at all. Now suppose she says something different. His customers are office staff buying on the way in, so the new office block opening two streets away will push his morning trade up, and if the office block fills and his morning takings do not move by the end of the coming year, she is wrong. The second sentence is a different kind of object entirely. The second sentence can be disagreed with. It can be watched. Twelve months from now, it can be settled.

A Research Thesis is the second kind of sentence, applied to a business that can actually be studied. One claim about one company, written in a single sentence, in language a colleague could push back on. The plainest test there is: if nobody could disagree with the sentence, it is a description of the company rather than a thesis about it. Sentences like the company is well run, the sector has good long term prospects, or management is focused on execution all fail that test instantly. Nobody is going to argue. Nothing is at stake. Nothing can be checked.

The distinction is worth sitting with. Most of what circulates under the word thesis is description. Description is not worthless. A reader who does not know what a company sells needs it. But description cannot be wrong, and a claim that cannot be wrong cannot teach its author anything. Teaching its author something is the whole reason a research process exists.

What four things does a thesis have to carry?

A thesis carries four parts, and each one is doing a job the other three cannot do. The claim is the thing being asserted. The variables are the small set of quantities that decide whether the claim comes true. The breaking evidence is the observation agreed, in advance, to show the claim wrong. The period is the stretch of time over which the whole apparatus can be judged.

Now the useful part. Dropping any one of the four does not produce a slightly weaker thesis. Dropping one produces a specific, recognisable object with its own name, and all four of those objects are common in published writing. A claim with no variables is a slogan, a claim with no breaking evidence is a preference, a claim with no period can never be wrong, and a set of variables with no claim on top is a spreadsheet.

How to build a testable research thesis, and why does the order decide everything?

The four parts are not written down in the order one might guess, and the order is the whole trick. The four parts form a sequence rather than a list. A list invites the parts to be done in whatever order feels comfortable, and comfortable is exactly what produces an untestable thesis.

Step one, start from something in the record that is genuinely unresolved. Not from a company that happens to appeal. Sarvani Coatings Limited has a real open question sitting in its published statements: gross margin rose from 43.0 per cent in year one to 46.0 per cent in year three, a gain of about 3.0 points across those two years, of which about 2.0 points arrived in the single step from year two to year three. The statements do not say why, and three different explanations fit the same numbers. An unexplained margin gain is worth having a view about. A view there could be right or wrong.

Step two, write the claim in one sentence. If it takes a paragraph, there is more than one claim present and one of them has to be chosen. Step three, strip the variables back until only the ones that decide the answer are left. Most models have forty inputs. A thesis usually rests on two or three, and the discipline is subtraction, not addition.

Step four, write the disconfirming evidence, and write it now. Before any argument in favour has been assembled. Step four is the step everybody moves, and moving it is what quietly destroys the thesis. Writing the breaking evidence at step four, rather than last after the case in favour is already built, is the single change that separates a testable thesis from an argued one. Evidence chosen after a view is held gets chosen, without any dishonesty at all, to be evidence the view can survive. The observations picked are unlikely, or vague, or so far away that they cost nothing. Written first, before any attachment has formed, the same person writes something much harder.

Step five, fix the period. Five years, three years, four quarters, whatever the claim actually needs, but a stated one. Step six, and only now, look at whether the price already contains the claim. Doing that first is how a view gets reverse engineered out of a price, and a view built backwards out of a price will always look supported.

What does the whole worked thesis look like on one sheet?

Here is the exercise, built in that order. Read it as an artefact, not as a view. The claim first, in one sentence: Sarvani Coatings Limited's gross margin gain is a level shiftA one time step up to a new resting point, as opposed to a rate that keeps repeating year after year. A cook who moves to a bigger stall once has taken a level shift; a cook whose queue grows every month has a rate. rather than a durable rate, so the earnings growth the price appears to assume will not be delivered from margin.

Notice what that sentence does. The claim is not saying the margin will collapse. It is saying the margin found a new level and will sit there. A margin sitting at a new level is a much more specific and much more checkable claim than a general worry. Somebody could flatly disagree, and it is clear what they would be disagreeing with.

The variables, stripped to three. Whether gross margin holds at 46.0 per cent. Whether volume growth stays above the sector's 4.5 per cent. Whether the 0.13 percentage point share gain repeats. Everything else in a model of this company is detail. The breaking evidence, written before any case in favour: gross margin holding at 46.0 per cent across four consecutive quarters while input cost per unit rises, or a share gain of 0.13 percentage points or more repeating in the following year. Either would show the claim was wrong. And the period: five years, the stretch the price arithmetic below runs over.

Where does the gap of about 15.0 percentage points actually come from?

Now the numbers, and every one of them can be worked with a pen. The illustrative price is Rs 486/- a share. The published year three result was Rs 278 crore of profit after tax. On 24.00 crore shares that is Rs 11.5833/- a share, printed as Rs 11.58/-. Dividing the price by the published figure puts the shares at about 42.0 times. Dividing by the unrounded figure instead gives 41.96 rather than 41.97, a difference that changes nothing anywhere below to one decimal place. The rounding check is worth doing once to establish that rounding does not matter at this price. In plenty of other places it does.

The next two numbers belong to the analyst, not to the market, and they are marked as such at every point where they are used. Take 12 per cent a year as the required returnThe annual return an investor requires before parting with the money. It is a chosen input, not a figure the market publishes, and two people looking at the same share can honestly use different ones., and take the view that the shares would change hands in five years on 25 times earnings. Then Rs 486/- has to become Rs 856.50/- over those five years. Running Rs 486/- forward at 12 per cent a year, compoundingGrowth applied on top of last period's result rather than on the original amount, so each year works on a slightly larger base than the one before it. as it goes, lands on Rs 856.50/-. At 25 times, reaching Rs 856.50/- needs earnings per share of Rs 34.26/-. And growing Rs 11.58/- into Rs 34.26/- over five years needs about 24.2 per cent a year. The 24.2 per cent is not a forecast anybody published; it is the implied growthThe earnings growth rate a price already contains, once a required return and an exit multiple have been fixed. It is arithmetic run backwards from a price, not a prediction anybody made. the analyst's own two assumptions put inside the price.

Against that, the thesis case. Hold volume at 6.0 per cent and hold realisationRevenue per unit sold. If the same quantity went out of the gate and revenue rose, realisation rose. It moves with price and with what mix of products was sold. at 3.0 per cent, and revenue grows at about 9.2 per cent. The 9.2 per cent works as 1.06 times 1.03, or 9.18 per cent, not 6.0 plus 3.0. The two rates multiply, and the habit of adding them costs more than it looks worth on a five year compounding.

Now the bridge, and it has to be said out loud rather than left implied. Left implied, the subtraction below looks like it is comparing two different things. The 24.2 per cent is an earnings growth rate and the 9.2 per cent is a revenue growth rate, and they become comparable only because the claim itself asserts that margin holds flat at its new level. If margin holds and the costs below it move with revenue and the share count and tax rate do not change, then earnings grow at the same rate as revenue. Equal growth is an assumption of the exercise, stated openly so that it can be attacked, and it is precisely the assumption the level shift claim is making. Subtract, and the gap is about 15.0 percentage points a year.

The two paths in the drawing settle the argument about time. After one year the two paths differ by Rs 1.74/- a share. Nobody would notice a difference that size in a quarterly result. By year five they are Rs 16.30/- apart, and one is 1.91 times the other. A thesis that is right about the direction and silent about the period is untestable for exactly this reason: over one year both assumption sets look almost identical, so any single year of data is consistent with both. The period is not decoration. The period is what makes the claim capable of being settled at all.

Thesis Change: what does it take to change the claim itself?

A Thesis Change is a change to the claim, and it happens for exactly one reason: one of the variables the claim depends on behaved differently from the way the claim required. Not because the share price moved. Not because a competitor said something on a call. Not because a good article turned up. Because a named variable, one of those written down at step three, did something the claim said it would not.

Here is the working benefit of having stripped the variables back, and it is bigger than it sounds. If the claim rests on three variables, then only three kinds of news can change it, and everything else that arrives is noise dressed as information. On this worked thesis, the three are gross margin holding at 46.0 per cent, volume growth against the sector's 4.5 per cent, and whether the 0.13 percentage point share gain repeats. A new plant announcement is not one of them. A change of chief financial officer is not one of them. A broker note is not one of them. All three might be interesting. None of the three can change this claim. This claim never rested on any of them.

The three variable limit is also why a genuine thesis change is rare by construction. Because a thesis change is an event, it should feel like one, and it should be traceable back to a specific variable named in advance. Where no variable can be pointed at, what changed was a mood.

Estimate Revision: what actually moves when a number is replaced?

An Estimate Revision is a change to a number inside the model. Most often it happens for a completely mechanical reason: a period got published, and the forecast being carried for that period was replaced by what actually turned up. Nobody changed their mind about anything. A cell that held a guess now holds a fact.

Work it on the record. Before the year was published, the consensusThe average of the estimates several people publish for the same line. It describes what other people are expecting, and it measures nothing at all about the business itself., being nine estimates for earnings per share, stood at Rs 11.17/-. On that base, the same illustrative Rs 486/- was 43.5 times, and running the identical arithmetic as before, an implied growth of about 25.1 per cent a year. Then the year was published at Rs 11.58/-. Now the same Rs 486/- is about 42.0 times, and the implied growth is about 24.2 per cent.

Two details matter here more than the numbers do. First, the target the arithmetic runs to, Rs 34.26/- of earnings per share in five years, did not move at all. The target depends only on the price, the required return and the exit multiple, and the publication touched none of the three. Only the starting point moved. Second, the fall of about 0.9 of a percentage point is computed by running the growth arithmetic twice, once from Rs 11.17/- and once from Rs 11.58/-. The fall is never computed by subtracting a printed 24.2 from a printed 25.1. Two rounded percentages subtracted give a difference with the rounding of both baked into it. Work differences in the underlying amounts and state the difference afterwards.

Thesis Change vs Estimate Revision: which single question separates them?

The distinction between a thesis change and an estimate revision is the sharpest of them all, and it comes down to a single question asked before anything is touched. Which of the named variables moved? If none of them did, the number changed and the view did not, and what is in hand is a revision rather than a thesis change.

Run it on what just happened. The year was published at Rs 11.58/- against nine estimates averaging Rs 11.17/-. Did gross margin stop holding at 46.0 per cent? No, the published gross margin is 46.0 per cent. Did volume growth fall below the sector's 4.5 per cent? No, volume grew 6.0 per cent. Did the share gain fail to repeat? Nothing about a following year is in this publication at all. None of the three moved. Therefore the arithmetic around the view moved and the view stayed exactly where it was, and the honest note to write is one sentence long.

Now the other direction. Gross margin holding at 46.0 per cent through four consecutive quarters while input cost per unit rises is a thesis change, and a clean one. A named variable would be behaving in the way the claim said it would not. Very little drama is required for it. No crisis, no scandal, no surprise. Just a variable named in advance doing the other thing.

Confusing the two costs something specific, and it costs in both directions. Treating every revision as a thesis change moves the view weekly, so it can never be tested; treating a genuine thesis change as a revision keeps a broken claim alive underneath a repaired number. The first failure is loud and looks like diligence. The second is quiet and looks like consistency. The second is worse.

Catalyst vs Thesis: why is a dated event not evidence about a business?

A catalyst is a dated event that could cause other people to revise their assumptions. A thesis is a claim about a business. A catalyst and a thesis are two different kinds of object, and the difference is not one of degree: a thesis is about a company and a catalyst is about an audience.

The worked thesis has three catalysts hanging off it. Quarterly results. The coatings line sitting in capital work in progressAn asset the company is still building. The spending sits on the balance sheet, the asset is not yet running, and nothing about it reaches the profit ladder until the day it is commissioned. at Rs 118 crore being commissioned. And any change to segment disclosureThe split of revenue and profit by line of business that a company chooses to publish. What sits inside each segment is set by the company, so the shape of the split can change between years.. Each of the three is a real, dateable event that would make people look. And now the move that has to be repeated every single time catalysts appear anywhere: take all three straight back out of the evidence. Not one of the three is information about whether gross margin holds at 46.0 per cent, and gross margin holding is what the claim actually rests on.

The commissioning is the clearest case. The coatings line will be commissioned or it will not, and the commissioning is a genuine event with a genuine date that might well move the price on the day. The commissioning says nothing whatever about whether a margin gain was a level shift or a durable rate. Whether an event occurs and whether a claim is right are simply different questions, and a catalyst answers only the first.

Which leads to the caveat that survives everything else in this guide. A thesis that needs a catalyst before anyone will recognise it is a claim about other people rather than a claim about a company. A case that only works when something makes the crowd look is not a view about a business, it is a prediction about attention. The expectation gapThe distance between what a price appears to assume and what the analyst's own case assumes. It measures a disagreement between two sets of assumptions and is taken at full depth separately. and how catalysts relate to it are taken at full depth in their own material. In a thesis, the only job is to keep catalysts out of the evidence.

What gets called a thesis and is not one?

Four things get called a thesis and are not one. A target for a share price is not a thesis. A rating is not a thesis. A story about a management team is not a thesis. A summary of what a company does is not a thesis. A target and a rating are the two that get published most often, and the reason both fail was settled in the valuation work: a single number carries no claim, no variables, no breaking evidence and no period, so there is nothing in it for anybody to test.

The management story deserves a moment because it is the most seductive of the four. Stories about people are memorable, they explain things satisfyingly, and they are almost impossible to disconfirm. What observation would show that a management team is not focused on execution? Where no quick answer comes, the sentence is not carrying a claim, and it will absorb whatever happens next in either direction.

What does a working thesis look like from the outside?

A good thesis is usually recognisable in about thirty seconds, without any knowledge of the company. It is short. The variables are named, and there are two or three of them rather than eleven. Its breaking evidence was written before any argument in favour. Evidence written in that order is specific and uncomfortable rather than remote. And it can be shown to be wrong on a date.

A thesis nobody could ever falsify is not a strong thesis, it is a claim that costs its author nothing. The point is worth saying plainly. An instinct runs the other way, an instinct that a view which survives all objections must be a robust one. The reverse is true. A view that survives all objections usually survives them because it was never exposed to any. The strength of a thesis is measured by how cheaply it could be destroyed, and by whether its author wrote down the instrument of destruction in advance and left it lying on the table.