How Estimate Revisions Can Change Research Context

An estimate revision changes the number a view is measured against, not the view. When the earnings base for Sarvani Coatings Limited, an invented coatings maker, moved from an estimated Rs 11.17/- a share to a published Rs 11.58/-, the same illustrative price of Rs 486/- went from 43.51 times to 41.97 times, and the growth it appears to assume fell from 25.13 per cent to 24.23 per cent. Nothing traded.

Underneath that sit one division and one fifth root. A multipleA share price divided by an earnings figure for the same share. The division is the same whichever earnings figure goes underneath. is a price over an earnings figure, and the growth a price appears to assume is that same earnings figure carried forward to a rating and a return the analyst picked. Replacing the earnings figure moves both readings on their own, with the market shut and the notebook untouched. Replacing one figure is the whole of what a revision does to the arithmetic, and it is exactly why a revised number and a revised opinion have to be kept in separate columns.

What is an estimate revision, and what does it leave completely alone?

Start with a kitchen table rather than a model. A household planning a wedding writes down Rs 4,00,000/- for the caterer. The last one cost that much, and nobody has a better number yet. Three weeks later the caterer sends the actual bill, Rs 4,15,000/-. The household replaces the guess with the bill. Nothing about the wedding changed. The date did not move, the guest list did not move, the reason for holding the wedding at all did not move. One cell that was always a guess stopped being a guess.

Replacing the guess with the bill is an estimate revision, and it is the most common event in equity research by a wide margin. A period closed, a company filed, and a figure that was somebody's forecast became a published fact. Before Sarvani Coatings Limited published year three, nine estimates of its profit after taxWhat is left of a year of trading after every cost, every interest charge and the tax bill. Profit after tax sits at the bottom of the profit ladder and is built up in the accounting material. existed, with a consensus meanThe plain average of the estimates a data compiler has collected for one figure from several analysts. The mean is an average, not a forecast anybody actually made. of Rs 268 crore and a spread from Rs 255 crore to Rs 284 crore. Afterwards there was one number, Rs 278 crore, and the nine forecasts were history.

A revision replaces a number inside a model and touches nothing else: not the claim, not the variables the claim depends on, not the evidence that would break it, and not the period over which it could be judged. Hold on to that list, because everything difficult about this subject comes from people quietly treating a replaced number as though one of those four things had moved with it.

How can a share get cheaper on a screen when nobody has bought or sold anything?

Look at the division rather than the market. The price is Rs 486/-, illustrative and as at a single stated date, and it sits on top. Underneath it goes an earnings per shareThe year's profit spread over every share in issue. Dividing that way lets a figure in rupees and paise sit beside a price. figure. The day before the filing that figure was the mean of the nine estimates, Rs 11.17/-, and Rs 486/- divided by Rs 11.17/- is 43.51 times. The day after the filing the figure was the published Rs 11.58/-, and Rs 486/- divided by Rs 11.58/- is 41.97 times.

Notice what did the work there. The numerator is identical in both panels. Every part of the fall sits in the denominator, and the denominator moved because a company filed a document, not because anybody formed an opinion. A share can become measurably cheaper or dearer on a screen with no transaction of any kind taking place, and that movement is arithmetic rather than a market event.

Precision matters in reporting the size of the fall. Worked from the rupee absolutes the record actually publishes, Rs 268 crore against Rs 278 crore, the multiple falls 3.60 per cent. Worked from the two rounded per share figures alone, Rs 11.17/- and Rs 11.58/-, it comes out at 3.54 per cent. The gap is pure rounding, it is small, and a note that writes 3.5 because it subtracted two printed decimals has already thrown away the precision the recomputation was for.

The share count, which is an assumption and is checked in the open

Every per share figure here runs through a share count, and the record publishes no share count. Twenty four crore shares is assumed, and an assumption that quietly drives four figures deserves a check rather than a footnote. The check is to run the assumed count forward against the two published per share figures it has to reproduce, Rs 11.58/- and Rs 61.92/-.

If the count were 24.5 crore, the same profit would give Rs 11.35/- and the same net worth Rs 60.65/-, neither of which is what the record prints. A forward check does not prove the assumption. The check shows that the assumption is at least consistent with two figures that were published independently of each other.

Why does the growth a price appears to assume fall when earnings come in higher?

The second reading works the same way and surprises people more. Take a price, decide the annual required returnThe yearly return an investor insists on before putting money into a share. The required return is a declared choice, and no filing anywhere contains it. sought from it, decide the exit multipleThe multiple a share is assumed to trade on at the end of the period being looked out to. Like the required return, it is a declared choice rather than a fact found in a filing. assumed at the end, and the arithmetic hands back the earnings a share the company would have to reach. The remaining question is what annual rate gets today's earnings there. The method itself belongs to the valuation material and is used here rather than rebuilt.

Moving the base is the whole mechanism, and the everyday version is a running track. Reaching a fixed line in five strides from a little further forward means each stride can be a little shorter. The earnings figure is the base of the growth calculation, so revising it upward pushes the implied rate down, and revising it downward pushes the implied rate up.

Then read the size of the fall for what it actually measures. Nought point nine of a point is the distance between what nine people expected and what the company reported, converted into the language of the price. The size of the move measures how far the estimate sat from the outcome, and so describes the forecast rather than the business. It is tempting to read a big move as the company having done something dramatic. A big move is at least as often the forecasters having been some distance out.

What does the whole published spread of estimates do to one unchanged price?

Now widen the lens. The mean of Rs 268 crore was never a forecast anybody made; it is the average of nine that were. The lowest was Rs 255 crore and the highest Rs 284 crore. Divided by the assumed twenty four crore shares, those give bases of Rs 10.63/- and Rs 11.83/-, and the unchanged price of Rs 486/- sits at 45.72 times against one and 41.08 times against the other.

Before a single share traded, nine ordinary bases already placed the same unchanged price across 4.64 times and 2.67 percentage points of implied growth. Most published revisions never produce that much movement. The spread is the calibration worth carrying away. When somebody says a multiple moved, the first question is not what happened in the market. The question to ask is which denominator they were using last week.

How is a revised number told apart from a revised opinion?

The sequence has already done the hard work here. A claim worth holding names the handful of variables it depends on before anybody argues about it. The worked exercise here rests on three: whether the 46.0 per cent gross marginGross profit written as a share of revenue, so what is left after the cost of materials and nothing else. Built in the accounting material and used here. stays put, whether volume keeps running ahead of a field growing 4.5 per cent, and whether another 0.13 percentage points of share is taken in the year after. So the test is a single question asked against a short written list, and it is answered by looking rather than by judging.

Two things follow, and the second is the one that gets people uncomfortable. The first is that most events are revisions. The second is that saying so is not laziness and it is not dismissiveness. Most published numbers are bookkeeping against a view. A process that treats every published number as news will change its view every few weeks, and leaves behind nothing steady enough to score. A view that survived four quarters unchanged is a view somebody can be held to. A view that moved eleven times is a weather report.

Why do revisions arrive in a bunch, and what does the bunch actually show?

Think of a street of tailors who all buy cloth from the same wholesaler. When the wholesaler raises his price on a Tuesday morning, every tailor on the street raises his price that week. Nine identical price rises look like the whole street reaching the same conclusion. It is one event reaching nine doors.

Research works the same way and for a duller reason. A results filingThe document a listed issuer lodges with the exchanges once a reporting period closes. Everybody receives the same one at the same moment. is lodged once, reaches everybody at once, and lands in nine models that all have roughly the same shape because they were all built to forecast the same lines. Revenue, margin, tax, profit, earnings a share. Feed the same new number into nine similarly shaped models and nine similar revisions come out within days of each other.

A cluster of revisions is evidence of a shared calendar rather than of shared conviction. Reading it as confirmation costs the analyst independence, the one thing the work was meant to buy. The cost is specific. Taking the cluster as agreement counts one fact nine times, and produces the greatest confidence precisely at the moment everybody is holding the same position for the same reason.

What should happen to an existing view when a revision lands?

Make it a procedure so that it does not become a mood. Four steps, in order, and the last one is the only one that is ever skipped.

Step four runs counter to how anybody feels after two hours of work, so it deserves a sentence of its own. A model was updated, two readings were recomputed, three variables were checked, and the honest output is that the view is exactly where it was. An unchanged view feels like nothing. Recording no change is the correct output most of the time, and it is worth writing precisely because it is unglamorous: a log that only contains changes cannot show later how steady the view was, only that it moved.

How does somebody doing this for a living actually use the log?

Three readers use the same four lines differently, and until each of them has been watched reading it back, the log looks like nothing except paperwork.

A research supervisor reads the log backwards at review time. Not to see whether the view was right. The market settles that on its own. The supervisor is checking whether the view moved for reasons written down in advance. An analyst whose log shows three revisions, three unchanged views and one dated change of mind against a named observation has demonstrably run a process. An analyst whose view drifts every quarter with no named trigger has demonstrably run a mood, however good the returns look.

A person managing money reads it as position sizing. If the update says the arithmetic moved and the claim did not, that is not a reason to trade, and the cost of trading on it is real money in charges and spread. If the update says a named variable moved in the breaking direction, that is a reason to act, and the size of the position was set against exactly that risk.

A household holding a few shares directly gets the most from the plainest version. One line is written at the time of buying: what has to keep being true. Then, each time the company files, the only question is whether that line is still true. Most times it is, the date goes beside it and the file closes, four minutes spent instead of an afternoon reading commentary. The whole practitioner value of the log is that it converts a stream of arriving numbers into a small number of dated yes or no answers against something written before its author had any stake in the answer.

What did the published year actually do to the numbers, and what did it do to the claim?

Run the exercise end to end. Before the filing there were nine estimates, a mean of Rs 268 crore and a spread from Rs 255 crore to Rs 284 crore. After it there was one figure: Rs 278 crore of profit after tax, and Rs 11.58/- a share. Against the mean that is a beat of 3.73 per cent, and it landed inside a spread that already existed. A beat that lands inside an existing spread is still usually reported as a surprise.

Now the test, taken one variable at a time against the published year. Gross margin came in at 46.0 per cent. The claim said that level would not hold as a rate, so the margin has not yet failed in the breaking direction. Volume grew 6.0 per cent against the field's 4.5 per cent, a one year figure against a one year figure. Volume held. Share moved from 4.87 to 5.00 per cent, a gain of 0.13 percentage points in one year. Share held too. None of the three moved in the direction the claim named as fatal to it. The correct entry in the record for this filing is that the numbers around the view moved measurably and the view did not move at all.

When is a revision genuinely a change of mind?

A genuine change of mind happens, and it is rare by construction. The named list is short and was written before anybody could observe the answer. The condition is narrow: the revision has to be to one of the named variables, and it has to run in the direction the claim said would not occur. Nothing else qualifies, however large the number that moved.

Hold the counterfactual beside the actual so the distinction has both sides. The claim under test says the margin improvement was a one time step up and not a rate that keeps running. Had the identical filing shown input cost per unit climbing over four quarters in a row while that 46.0 per cent refused to give way, the same document, on the same day, from the same company, would have been a change of mind. Not because the figure was bigger, but because that exact observation was written down in advance as the one that breaks the claim.

One limitation belongs in the open, and it is the kind of thing worth noticing whenever a breaking test is written. The invented record behind this exercise publishes revenue, earnings before interest, tax, depreciation and amortisation (EBITDA) and profit after tax by quarter, and it publishes no quarterly cost of materials. So it carries no quarterly gross margin at all. The breaking observation, as written, needs a disclosure that does not currently exist. The test as drafted cannot be run on what is filed today. A breaking test that requires a disclosure nobody publishes is not a strict test, it is an unrunnable one, and the fix is to rewrite the test against something that actually gets filed rather than to quietly stop checking it.

And when a genuine change of mind does arrive, the response is already settled: say the claim was wrong, in writing, beside the claim, with the date and the observation that broke it. Not a quiet edit to a spreadsheet. Beside the claim, where anybody reading the claim will also read the correction.

What happens when the revision runs towards the price instead?

One failure hides inside the procedure, and it is easy to commit while feeling diligent. The output sits a long way from the price. The cheapest repair available is not to rethink anything; it is to nudge one assumption until the two meet. Pushing the exit multiple from 25 times to 31 times closes the gap without a single line of the analysis changing.

The tell is not that an assumption changed but which assumption changed. Nobody can contradict the input sitting furthest from published evidence, so that input is always the one that moves. Volume growth and gross margin are in the filing and would embarrass whoever moved them. The exit multiple and the required return belong to the analyst, and moving them costs nothing except the entire point of the exercise. Once the model has been made to agree with the price, it can no longer say anything about the price, and every number in it is now decoration.