How to Run an Equity Research Post Mortem, Step by Step

A post mortem reads the work rather than the outcome. The call is retrieved as it was written, what happened is separated from what was claimed, the mechanism is recomputed, and then one assumption at a time is corrected until the error appears. The finding is then classified as a mechanism error, a magnitude error, an evidence error or no error at all, and written as a change to the process.

What is a post mortem actually examining?

A document. Not a result, not a share price, and not a person. The distinction sounds obvious right up until an analyst sits down with a call that went wrong and discovers that the outcome is already sitting on the desk, shouting, and that every judgement about to be made will be made in its hearing.

Here is the everyday version. A household decides to buy a scooter on instalments rather than take the bus, on the grounds that the monthly cost is similar and the time saved is worth something. Six months later the scooter is stolen. Was the decision wrong? Almost everybody says yes, and almost everybody is answering a different question from the one asked. The decision is examined by asking what was known, what was assumed, and whether the arithmetic held. The theft answers none of that.

A post mortem asks whether the reasoning that produced a conclusion was sound, and it can only answer that if the answer to a different question, namely what happened, is kept out of the room until the end. The separation between work and outcome is the whole architecture of what follows, and every step in the procedure exists to hold it in place under pressure.

Each of the next three claims was settled elsewhere, so it is assumed here rather than argued. A research error and a market outcome are different animals. A view has an assumption setThe short list of things a published view actually depends on, written down before the view goes out. How one is drawn up and kept current belongs to the thesis material., written down in advance, and that advance list is what makes any of this retrievable at all. And a published research document names the disconfirming testSomething written down beforehand which, if it were observed, would show the view is wrong. It is used here and taught in the thesis material. that would break it. Where a call carries none of those, the procedure below still runs, but step three does far more work.

What are the ten steps, and why that order?

Order is not decoration here. Steps three to six can only be done honestly while step two is sealed, and step seven cannot be done at all until they are finished. The shape of the list is the point, so it repays one reading as a whole before the worked instance.

Something is missing from that list. No step decides whether the call should have been made, no step grades anybody, and no step produces a view about the company. A post mortem is an examination of a document, and it ends in an instruction rather than a verdict about a person.

Step one: what did the call actually say?

Take the call in the teaching record. A wrong call is worth keeping rather than being embarrassed about. At the end of year two, with the year two figures published and gross marginRevenue less the cost of materials, expressed as a share of revenue. How the line is built and where it sits on a profit ladder belongs to the accounting material. standing at 44.0 per cent, an analyst wrote that Sarvani Coatings Limited, an invented maker of paints and coatings, would see its gross margin compress. Input cost per unit was rising and the company had limited room to raise prices.

The call is the object. One sentence, one date, one document, and everything that follows is an examination of it. Step one has already done a great deal of retrieval work: the date matters because it fixes what the analyst could have known, and the margin standing in front of them on that day was 44.0 per cent and not some other figure. A wrong date grades the call against a base it never used.

A remembered call is always a better call than the one that was written. Step one therefore insists on the document rather than the recollection. Memory quietly adds the caveats that would have made the call defensible, and it does this to everybody, including the reviewer who is certain they are being fair.

Step two: what happened, and where does it go?

Gross margin rose. The figure went from 44.0 per cent to 46.0 per cent over the following year, a gain of 2.0 points. The gain is written down and set aside, unused until step seven, and the temptation to look at it early is the strongest force in this whole procedure.

The habit being resisted has a name. Judging the quality of a decision by how it turned out is resultingGrading a decision by its result rather than by the reasoning that produced it. Named here only so the habit is nameable; where it comes from and what it does to a research process is settled in the sector material.. The term belongs to Annie Duke, Thinking in Bets, 2018. Knowing that margin rose makes every assumption in the call look either obviously right or obviously foolish, and neither impression survives contact with the arithmetic. The separation therefore has to be physical.

Step three: how is an assumption the call never wrote down recovered?

List what the call contained. The call contained one written assumption: input cost per unit would rise, at 3.6 per cent, to the one decimal place an analyst writes an assumption in. The call also carried a conclusion, compression, and compression cannot follow from that assumption alone.

Something else was needed and never appeared. For margin to compress, realisationRevenue per unit sold. Revenue divided by volume, so it moves when price moves and when the mix of what is being sold moves. per unit had to be assumed to grow more slowly than input cost per unit. The document does not say by how much. The missing figure is recovered from the conclusion instead: any figure below 3.6 per cent produces compression, so the recovered assumption is set at 3.0 per cent, marked as recovered every time it appears, and treated as evidence about the work rather than as something the analyst said.

A number the conclusion required and the document never stated is not a gap in the reconstruction, it is a finding about the work, and on this call it turns out to be the entire finding.

Step four: did the chain the call described actually operate?

One question, and no other. The call described a causal chain: input cost per unit rises, prices do not keep pace, margin falls. Take the first link and check it against what was published.

The cost of materialsThe single line covering the inputs that go into what is made. In this record it is by a wide distance the largest line on the ladder, which is why the whole margin story runs through it. rose from Rs 1,187 crore to Rs 1,304 crore, a rise of 9.86 per cent. Volume rose 6.0 per cent. So input cost per unit of output rose by 1.0986 over 1.06, or 3.64 per cent. The analyst wrote 3.6 per cent. The chain the call described operated, and it operated at almost exactly the rate the call put on it.

Stop there. Two conclusions are very tempting at this point: that the call was therefore fine, and that it was wrong anyway because margin rose. Both are answers to step seven, and step seven has not happened yet. All step four establishes is a yes.

Step five: how is the assumption that failed isolated?

Step five does the real work, and it is mechanical. Hold every assumption at the value the call used except one. Move that one to what arrived. Record how far the answer moves. Then put it back and do the next one.

The answer here is gross margin, and it is built from one relationship the reader already has: the materials share of revenue moves with input cost per unit and against realisation per unit. The call started from a materials share of 55.99 per cent, or Rs 1,187 crore over Rs 2,120 crore.

Read the last row first. The last row is what makes the other three trustworthy. Correcting both assumptions lands on 46.00 per cent, the published outcome. The reconstruction closes. Had it landed anywhere else, every figure above it would have been suspect and the honest move would have been to go back to step three.

Correcting the assumption the analyst wrote down moves the answer by 0.02 of a point, and correcting the one they never wrote down moves it 2.34 points, so a single unstated number carries the whole of the error.

Step six: how large is the error, in the units the decision was made in?

The gap between the call as published and what arrived is 46.00 less 43.68, or 2.32 points of gross margin. Nobody has ever felt 2.32 points of anything, so on its own that gap is a number and not yet a size.

So put it in three units. In points, 2.32. In rupees, at the revenue that actually arrived of Rs 2,415 crore, it is Rs 56.05 crore of gross profit. And as a share of the line it belongs to, it is Rs 56.05 crore against a gross profit of Rs 1,111 crore, or 5.04 per cent. Roughly five per cent of the line, and it reversed the direction of the conclusion. Either fact alone misleads, so the two are always reported together.

One caution about scales. Slipping between the two is easy. The call also implied a revenue growth assumption: volume at 6.0 per cent compounded with the recovered realisation of 3.0 per cent gives 9.18 per cent, against the 13.92 per cent that arrived, a difference of 4.74 points. The 4.74 points sit on the revenue scale. A revenue-scale figure is not added to the 2.32 points on the margin scale, and it is not subtracted from them either. Two scales, reported separately, and any arithmetic that mixes them produces a figure of nothing.

Step seven: what kind of error is it?

The sealed result comes back over. Margin rose 2.0 points. Set against what steps four to six established, the classification falls out without argument.

Work the ladder. Do the numbers do what the document says? Yes, the arithmetic in the call is sound. Was something checkable available and not checked? Hold that question. The answer becomes the finding at step ten, and no figure that existed on the day was skipped. Did the chain operate? Yes, established at step four. Was a size inside it wrong? Yes, and step five identified which one.

The verdict is a magnitude error. The failing assumption can be pointed at in the work today, so it is a research error rather than bad luck. Pointing at the failing number is the whole test for the second half. If the number that failed can be pointed at, together with what it should have been, it was research. If nothing in the document could have been different and the world simply went somewhere, it was an outcome.

Step eight: a narrow error, or a broken view?

Two questions again. How many of the assumptions failed, and did the chain survive? Here one of two failed and the chain held, so this is a narrow error. The scope of the finding follows from that: it applies to one assumption in one kind of call, and it does not touch how this analyst reads a business.

The mirror case looks identical from the outside, so it is worth naming. A view whose chain never operated is broken even if its final number came out close. Such a view was right by accident and will not be right the next time. From the result alone the two are indistinguishable. From the work they are not remotely alike.

Steps nine and ten: what does a post mortem actually produce?

An instruction, and a check. Nothing else counts as output.

The obvious process change here is wrong. Most reviewers reach for it anyway, so it is worth saying so plainly. The obvious change is to assume higher realisation next time. Raising the level repairs this one call and breaks the next. The missing item was not the level of the assumption. The missing item was the assumption.

The real change is that a call on margin must state its realisation assumption as a number, and state alongside it the level at which the conclusion reverses. On this call that level was computable on the day, from figures already on the analyst's desk, and it is not even a new number. Gross margin is unchanged exactly when realisation per unit and input cost per unit grow at the same rate, so the level at which the conclusion reverses is the input cost assumption itself, restated. The analyst assumed 3.6 per cent. The reversal level was therefore 3.6 per cent. The recovered realisation assumption sat 0.6 of a point below it, a thin margin for a directional call, and what arrived sat 3.87 points above it.

Step ten turns that into something a person can do. The check is: before a margin call goes out, the level at which it reverses is computed and printed beside the assumption. Where does that check sit? In the working order, immediately before the document is passed for review. A change is still cheap at that point and expensive after it.

When is the honest finding that nothing needs to change?

Sometimes step seven lands on no error. The chain operated, every size was defensible on what was known, the arithmetic held, nothing checkable was skipped, and the result went somewhere else anyway. Such a post mortem happens, and its finding is a real result.

Write it down explicitly: no error found, no change warranted, with the date and the call it refers to. A review that produces nothing feels like a review that was not done, so the temptation is to find something. Give in to that once a quarter and within two years the working order carries a stack of rules each of which would have prevented exactly one past error and none of which will prevent the next. A working order like that is ritual, and the only defence against it is being willing to write the word none.