Errors and Review: Classifying, Correcting, Disclosing and Checking

Errors in finance work come in four kinds: fact, method, presentation and omission. The response depends on whether the work has left the author's hands and whether the conclusion moves. A correction is silent before delivery, a correction notice follows delivery, and a reissue follows where the conclusion changes. A review pass exists to find these before the reader does.

One thing has to be settled before anything else. An error is not evidence that somebody was careless. An error is the ordinary output of a process in which a person carries numbers between documents, and any process built out of that will produce errors at a steady rate for as long as it runs. Treating errors as rare is what makes them shameful, and shame is the single largest cause of delayed disclosure. Treating them as expected is what turns checking from an accusation into a routine, and the routine is the only thing that reliably catches anything.

The same thing is familiar from somewhere non financial. A household that adds up a month of spending by hand will get a different total on the second attempt, and nobody in that household concludes that somebody is bad at arithmetic. The household adds it again. The adding again is not an insult, it is part of the adding. Finance work that goes out without the equivalent step has skipped part of the work, not part of the courtesy.

One assignment carries the craft from the first draft to the delivered note. The Kavery research desk, four people inside Kavery Capital Services Private Limited, an invented lender, was asked a question about Meenakshi Tubes Private Limited, an invented borrower the desk already lends to. Sharada Iyer produced the work. Prakash Nadar ran the checking pass on 11 March, before anything was delivered. Latha Menon commissioned it and recorded the decision on 12 March.

The borrower's product never enters the craft at all. The subject is what happens to finished work between the draft and the delivery, so a bond, a warehouse, a mutual fund or one machine in its place changes not a single sentence below. A checking habit that only works on one kind of thing is not a habit, it is a coincidence.

What counts as an error, and what is just a difference of judgment?

An error is a statement in the work that is wrong on its own terms. Somebody who has the same documents in front of them, applying the same stated method, would land somewhere else. Notice how narrow the test is. The test says nothing about whether the conclusion was wise, whether the assignment was worth doing, or whether a different analyst would have framed the question differently.

A difference of judgment is the opposite shape. Two people hold the same documents, agree on every figure in them, and still choose differently, because the choice was never determined by the documents. Splitting a cost total sixty and forty between the part that moves with revenue and the part that does not is a judgment. Nobody filed that split. Someone chose it, and someone else could choose seventy and thirty without either of them being wrong.

The practical difference is what fixing it looks like: an error is corrected, a judgment is disclosed. A judgment cannot be corrected. There is nothing to correct it against. A judgment can instead be written down as a choice, with the name of whoever made it and what happens to the answer if it goes the other way. Work that argues about its judgments in the correction log has an author who has not separated the two things.

The opposite mistake is more common and does more damage. Somebody finds a wrong number, and it gets discussed as though it were a matter of view: perhaps the other figure is defensible, perhaps it depends on the angle taken. It does not. A number computed on the wrong input has no defensible reading. Calling an error a judgment is the most comfortable mistake available at the moment of finding one, and it is the reason the four kinds below get named out loud rather than felt.

How to Handle Errors in Professional Finance Work: What Are the Four Kinds?

Four kinds cover everything that turns up in finished finance work, and they are worth naming separately for one reason only: they demand different responses and they are found by different checks. Naming the kind is therefore the first move rather than a label attached afterwards. Name it wrongly and the check that runs was never going to find it.

Kind one, fact: a number that is wrong, whatever the reason

An error of factA number that is wrong, whatever the reason it is wrong. is the plainest of the four. The work says one thing and the source says another. The route in does not matter, whether the number arrived by mistyping, by dragging a formula one cell too far, by reading the wrong column of a table, or by writing down what somebody said on a call. The kind is settled by the outcome, not by the route.

In this assignment it looked like this. The draft carried the operating margin for the year reported at 12.5 per cent. The 12.5 per cent comes from dividing operating profit of Rs 2,30,00,000 by revenue of Rs 18,40,00,000, and Rs 18,40,00,000 is the revenue of the prior year. The revenue of the year reported is Rs 21,20,00,000, and the same profit divided by that gives 10.8 per cent. One column too far to the left, in a spreadsheet where the two years sit next to each other.

Nothing about the arithmetic was wrong. An error of fact is hard to see from the inside for exactly that reason. The formula was the right formula. The division was performed correctly. A person rechecking the sum would confirm the sum. The only way to catch it is to stop looking at the calculation and start looking at where each input came from. Tracing inputs is a different activity with a different rhythm, and it is the first of the four checks below.

Kind two, method: the arithmetic is right and the wrong thing was computed

An error of methodThe arithmetic is right and the wrong thing was computed. is the one that survives almost every check there is. Every input traces cleanly to a source. Every figure is the figure that was filed. Every operation was performed correctly. And the quantity that came out is not the quantity the sentence claims it is.

The everyday version is familiar. A household wants to know how much the monthly rent has risen over three years, so it takes the three annual increase percentages, adds them and divides by three. The arithmetic is faultless. The number produced is the average of three increases. An average of three increases is not the total increase, and the two answers can be a long way apart. Nobody typed anything wrongly. The wrong question was answered accurately.

In finance work the commonest form is exactly that: three yearly margins averaged and then described as the three year margin. The three year margin is total profit over the whole period divided by total revenue over the whole period. Averaging the three ratios gives each year equal weight regardless of its size, so a small year with a fat margin pulls the answer up out of all proportion to how much business it represents. Nothing is wrong to find, so rechecking cannot surface a method error at all. Only somebody asking which quantity the sentence claims will surface one.

The pass in this assignment found no error of method, and the absence is worth naming. A checking habit learned only from the errors that were caught takes the shape of the catches, and that shape has a hole in it exactly where method errors sit. Had there been one here, no amount of tracing to the log would have surfaced it. The fourth check in the pass is therefore not a check on the numbers at all.

Kind three, presentation: the number is right and the reader will read it wrongly

An error of presentationThe number is right and the reader will take the wrong meaning. is the kind that gets skipped. Every instinct for finding errors is aimed at numbers, and in this kind the numbers are all correct. The wrong part is the container. An axis that starts somewhere other than zero without saying so, two series drawn on different scales in one frame, a percentage of a percentage described as a percentage, a colour that means good on one chart and bad on the next, a rounded figure sitting in a column of unrounded ones.

The draft in this assignment carried a chart of the operating margin for the two years, and its vertical axis began at 9 per cent with the top of the frame at 13. Both plotted values were right: 12.0 per cent for the prior year and 10.8 per cent for the year reported. The fall between them is 1.2 points. Drawn inside a frame only four points tall, that 1.2 point fall occupied 30.0 per cent of the whole picture and read, at a glance, as a collapse.

Redraw the same two points on an axis that starts at zero and the same fall occupies 9.2 per cent of the frame. The fall then looks like what it is: a real decline, worth the attention of the person who commissioned the work, and not a cliff. Nothing about the data changed between those two pictures. An error of presentation counts as an error rather than a matter of taste for that reason alone. A reader who takes the wrong meaning from a correct chart has been misinformed just as surely as by a wrong number, and they have no way of knowing it happened.

The panel below turns that one decision into something adjustable. The control is the only thing that changes: where the bottom of the vertical axis sits, anywhere from 0 up to 10 per cent. Both plotted values stay pinned at 12.0 and 10.8 per cent and the top of the frame stays pinned at 13 per cent, so the share of the frame that the 1.2 point fall occupies is simply 1.2 divided by 13 minus the axis start. At a start of 0 that is 9.2 per cent of the frame. At the default setting of 9, the setting the draft actually used, it is 30.0 per cent. At 10 it reaches 40.0 per cent, more than four times what an honest frame gives it.

Two things are worth noticing while the control is in use. The first is how early the distortion arrives: by an axis start of 5 the fall already fills 15.0 per cent of the frame, and 5 per cent is a setting that feels reasonable to anybody who is thinking about fitting the data neatly rather than about what the reader will see. The second is that no setting is dishonest by itself. A non zero axis is a legitimate choice when a chart exists to show a small movement in a large quantity, and the correction is not a ban but a label: where the axis starts, stated on the axis, in type the reader can read.

Kind four, omission: nothing present is wrong and something load bearing is missing

An error of omissionNothing present is wrong; something load bearing is missing. is invisible by construction. Every sentence in the document is accurate. Every figure traces. And the reader still ends up with a false picture. The fact that would have changed their reading is not there to be read.

Picture a caterer quoting for a wedding. Every line on the quote is correct: the per plate rate, the number of plates, the arithmetic at the bottom. The service charge and the transport are simply not mentioned. Nobody has stated anything untrue, and the person reading the quote has still formed a wrong idea of what the evening will cost. The missing charges are an omission, and the only way to find one is to ask what is supposed to be in the document rather than to check what is.

In this assignment the omission was small and precise. One sentence in the draft carried no item number. Every other claim in the work was tied to a numbered item in the source log, and this one was not. Traced back, it turned out to rest on the call with the finance head on 10 March rather than on anything filed. A call is a different weight of support entirely, and the reader had no way of telling. The check that finds an omission is not a check on any number but a check that the pattern is unbroken. A pattern check runs over the whole draft in one separate pass rather than figure by figure.

Which two questions decide what happens next?

Once something is found and its kind named, the response is decided by two questions and nothing else. Has the work left the author's hands? And does the conclusion move? The two questions are the entire test, and their shortness is the point: a person who has just found their own error is in no state to design a bespoke response, so the response is designed in advance.

The first question is about circulation, not about time. Work has left the author's hands the moment somebody outside the process could be acting on it, whether that happened three weeks ago or eleven minutes ago. Work sitting on the author's own screen has not left their hands even if it was finished yesterday. Work sent to one colleague for review has not left them. Work sent to the person who commissioned it has.

The second question is about the answer, not the number. The conclusion moves when the corrected figure changes what the reader would conclude. Usually that means the figure crosses a threshold, reverses a direction, or contradicts the sentence at the top of the note. A figure that shifts from 10.4 to 10.6 per cent under an 11 per cent floor has not moved the conclusion. A figure that shifts from 11.2 to 10.8 has, and the size of the two changes is almost the same.

Error Correction: When Is the Fix Silent, When Is It a Notice, and When Is It a Reissue?

Three routes come out of those two questions, and they differ in who is told rather than in how carefully the work is fixed. The fixing is identical in all three. Only the size of the circle that finds out changes.

A silent correctionFixing it before the work leaves the author's hands, with the change logged. is what happens while the work is still with its author. The figure is changed, one line goes into the working file saying what was changed, when and why, and the draft carries on. The word silent describes the outside world, not the file. A correction nobody can see later is indistinguishable from a number that was always there, so the change is recorded in full. All three findings in this assignment were handled this way, for the simple reason that Prakash Nadar ran the pass on 11 March before anything was delivered.

A correction noticeTelling the reader what changed, when, and what it does to the answer. is what happens once somebody outside could be acting on the work and the conclusion still stands. The notice goes to everybody who received the original, not to the person who happens to have asked. A notice is short, it is fixed in shape, and it does not explain itself. A notice that spends three paragraphs on how the error happened has converted a correction into a defence, and the reader now has to work out what actually changed.

A reissueReplacing the whole document because the conclusion itself moved. is what happens when the conclusion has moved. The document is replaced rather than annotated, the original is marked superseded, and the replacement carries the same item numbers so that anybody holding either version can line them up. The reason a notice will not do here is circulation: a note whose top line says one thing, with a separate message somewhere saying the opposite, leaves two contradictory records in the world, and the one that gets forwarded is always the document rather than the message about it.

Quality Review: What Does the Pass Check, and Why in That Order?

A review passA separate reading whose only job is to find what the writer cannot see. is not the writer reading the draft again. A review pass is a separate reading with a different starting point, and the difference in starting point is the entire mechanism. The writer reads the draft and tests whether it holds together. The pass reads the log and tests whether the draft matches it. The two readings find completely different things, and only one of them finds the error that is entirely convincing.

Four checks, always in this order. Each one assumes the one above it has passed. Running them in a different sequence is not slightly worse, it is largely wasted. Confirming that a claim carries an item number is worth nothing if the number in the claim was never traced. The result is a certificate that a wrong figure is properly sourced.

Check one is where the work is. The check starts at the log, not at the draft. Item one is opened, the figure is read off it, and the place where that figure appears in the draft is found. Then item two. Working in this direction rather than reading the draft and asking where each number came from matters more than it sounds: reading the draft first lets the draft dictate what to look for, and a draft that contains a confident wrong number will point at the wrong thing.

Check two is a pattern check and it goes quickly. Check two does not judge whether any source is any good. The check runs down the draft asking one question of every claim: is there an item number attached. Anything without one is pulled out and traced, and in this assignment exactly one sentence failed it.

Check three is the one people skip, and errors of presentation are the reason it exists. Every axis, every unit, every scale, every label. Where does each axis begin. Is a change in points being described as a percentage change. Is one chart in lakh sitting next to another in crore. None of these are number errors and all of them change what the reader takes away.

Check four puts the numbers down entirely. Cover the whole note and read only the answer at the top. Does it say what the work actually found. Is it a conclusion or a description of activity. Would somebody who reads nothing else be correctly informed. Only check four can catch an error of method. Only check four asks what the numbers are supposed to mean rather than whether they are right. The answer at the top of a note carries the load in this order because that is how a reader takes it in, a shape Minto set out in The Pyramid Principle in 1978.

How to Perform a Finance Work Error Review: What Does It Cost and What Does It Find?

The argument for a checking pass is almost always lost on time and almost never on principle, so here is the whole assignment measured in minutes. Defining the question took 45 minutes. Gathering took 4 hours. Computing took 2 hours. Drafting took 2 hours and 15 minutes. Production came to 9 hours. The pass took 40 minutes.

Forty minutes of checking against nine hours of production is 7.4 per cent of the building time, and 6.9 per cent of the 9 hours and 40 minutes the work took end to end. Those two figures sit on opposite sides of seven, and a sentence that is true against one base is false against the other, so a ratio like this one has to name which base it uses. Naming the base is the same discipline as naming the denominator under a margin, and errors of method above set out what happens when nobody does. Either way the pass is a rounding error in the effort, and it caught the one thing in the draft that would have reversed the answer. That ratio is worth carrying around, because the moment the pass gets cut is always the moment the work is running late, and running late is exactly when the draft contains the most transcription.

Two practical notes on running one. A gap between finishing and checking, even a short one, matters because the pass depends on not remembering what was meant. And the findings are written down as the pass goes rather than fixed as they are found, because fixing pulls the reviewer back into drafting, and drafting is the state of mind the pass exists to escape.

What did all four kinds look like on one assignment?

Here is the assignment assembled, so the four kinds sit next to each other on the same evidence. The measure is the operating margin, the period is the year just reported, and the internal floor the desk works to is 11 per cent.

Three findings, three different checks, and not one of them would have been found by the other two. That is the argument for running a fixed set of checks rather than reading carefully. Careful reading is a single instrument and it is tuned to plausibility. The pass is four instruments, and only one of them is tuned to plausibility at all. Because nothing had been delivered on 11 March, all three were corrected silently, logged in the working file, and the note went out on the corrected figures the same day.

How is work reviewed when there is nobody else to review it?

Most people doing careful work do not have a Prakash Nadar. One person, one file, one deadline. The pass still has to happen, and the honest position is that self review is weaker than review by somebody else and is still worth an enormous amount more than nothing. The weakness is specific and it can be attacked directly: reading one's own draft replays one's own reasoning, and reasoning that produced an error will reproduce it.

The replay can be broken in three ways. Time between writing and checking, ideally overnight and at minimum long enough for something else to have happened. A change of direction: starting from the log and working into the draft, exactly as somebody else would, because a stranger has no memory of writing it and neither has an author who starts where a stranger would start. And a change of order: the numbers are checked in the sequence the log holds them rather than the sequence they were written in, so that no figure arrives with the sentence built around it.

Two small mechanical helps. Reading the draft aloud, because the mouth catches a missing clause the eye completes silently, exactly as it does with a message about to be sent to somebody important. And checking figures against the source with the draft covered, saying the number out of the source first and only then uncovering the draft, so that the draft is being tested rather than confirmed. Every one of these is the same move in a different costume: stopping the draft from dictating what to look for.

Who outside the work actually uses any of this?

The correction record looks like housekeeping until somebody who was not there needs it, and then it is the only thing in the file doing any work. Four situations, all ordinary.

The person who inherits the file. Six months on, somebody else on the desk picks up the same borrower and needs to know whether the 10.8 per cent figure is settled or contested. A logged silent correction tells them in one line that the figure was checked, that it moved, and what it moved from. Without the line, they find two figures in the history and no way to tell which one won.

The reviewer inside the firm. Whoever is asked, later, whether the work was done properly is not going to reread nine hours of production. They will look for evidence that a pass happened, what it checked and what it found, and a pass that found three things is far stronger evidence of a working process than a pass that found none.

The reader who received a notice. A lender, an analyst or an investor holding a note that has been corrected once reads the next note from the same desk differently, and they read it better. The reader now knows the desk finds its own errors, and that is the only observable signal anybody outside has about whether a checking pass exists at all. A correction sent promptly buys more credibility than the clean run it interrupted.

And the household version, because this is not a professional invention. A bank statement carrying a wrong entry that gets reversed with a dated line explaining it is worth more than a statement that quietly shows the right balance a week later. In both cases the value is not the fix. The value is that the fix is visible, so the reader can tell the difference between a record that has been checked and one that has merely never been questioned.

The Kavery research desk, Kavery Capital Services Private Limited, Meenakshi Tubes Private Limited, Sharada Iyer, Prakash Nadar, Latha Menon and the 11 per cent margin floor are invented.
Educational material. Not advice on any investment, tax, budget or market position.