RecalcfxSpreadsheets, and the wrong answers that look rightAll 42 entries
fxtwo of these find most of it, and neither is where people start
ABCDEFGH
1

6.5 · part 6

Reviewing somebody else’s sheet, and in what order

Reading every formula is impossible and is what people attempt. There is an order that finds most of it in half an hour.

For a separate operational view of time, ownership and team activity, see Monitask's practical guide.

2

Established onExcel 365 and 2016 and later: go to special, constantsGoogle Sheets, web: formula view onlyBoth: reconciliation needs nothing from the software

Reading every formula in an inherited workbook is impossible and is what people attempt. There is an order that finds most of what matters in about half an hour.

One: the shape

How many sheets, which are hidden, what the tab order suggests, how large the file is. A workbook of a few visible rows occupying several megabytes is carrying something, which the entry on what travels with a file explains.

Two: the typed cells

Select all constants within the calculated regions. That produces the list of numbers somebody typed where a formula was expected, and it is where the answer usually is.

This step takes ten seconds and it is not the step anybody starts with.

Three: the formulas, column by column

Formula view, then down each calculated column looking for a cell that reads differently, which the previous entry covers.

Four: the links and the names

What does this file point at, and what names does it contain. Both answer the question of what it depends on that is not visible, and both lists are usually longer than the author expects.

Five: the checks

Does the workbook contain any, and are they independent in the sense of the entry on cross-footing. A model with no checks is a model whose author has not asked whether it is right; a model with checks that share a path is worse, because it has asked and answered wrongly.

Six: reconcile one figure

Take a single number from the output and reproduce it from an outside source: the source system's own report, a bank statement, a colleague's independent count.

One figure is enough to establish whether the model is in the right territory, and no amount of formula reading establishes that.

The three questions for the author

What is this for. What would you not trust it for. Which single cell, if it were wrong, would matter most.

The third one produces better information than an hour of reading, and authors answer it accurately, because they know.

1the shape of the file2the typed cells3the formulas, column by column4the links and the names5the checks, and whether they are independent6one figure against an outside sourcetwo of these find most of it, and neither is the one people start with
6.5Six review steps in order, from the shape of the workbook to reconciling one figure against a source outside it.

The rate at which to work

Slowly enough to understand each formula you read, which means reading few of them. A review that covers a whole workbook quickly has established nothing, and the honest output of a fast review is a list of the parts that were not examined.

Saying which parts were not looked at is more useful than an unqualified approval, and it is what a reviewer is usually asked to give instead.

What a review cannot establish

That the model is right. It can establish that the arithmetic is consistent, that the inputs are where they should be, and that the outputs reconcile against something external.

Whether the approach is the correct one for the question is a different conversation, requires knowing the domain, and is not what anybody means when they ask for a review of a spreadsheet.

Writing it down

The output of a review is a list: what was checked, what was found, what was not examined, and what the reviewer would change. On the documentation sheet, dated, so the next reviewer starts from it.

Reviews that live in a message are reviews that are performed again from nothing next year.

What we cannot verify

The selection and inspection features referred to here are documented by the vendors and appear in the entries they point to. The order proposed is a recommendation based on where defects are found in practice, not a result from a study, and reasonable reviewers order these steps differently.

What to do when the author has left

Which is the usual case, and changes the order: start with step six, reconciliation, because it establishes whether the file is producing sensible figures at all before anybody spends an afternoon reading it.

Where the output cannot be reconciled against anything, that is the finding, and it is more important than anything in the formulas.

The half hour and what it leaves

It leaves the semantic questions, which are the ones this part's entry on tools describes as unreachable by any automation. Those need somebody who understands the subject and take as long as they take.

Separating the two halves is the practical value of an ordered review: half an hour of mechanical checking makes the remaining time count.

What to say when asked to sign it off

Say what was checked and what was not, in writing. A reviewer who approves a workbook they read for half an hour has made a claim about the whole of it, and the claim they can actually support is narrower and more useful.

The one thing to do if there is no time at all

Select the constants inside the calculated regions and look at the list. Ten seconds, one dialogue, and it is where the answer is often enough that no shorter check is worth performing instead.

In short

  1. Reading every formula is impossible and is what people attempt.
  2. Selecting the typed constants in calculated regions takes ten seconds and usually finds it.
  3. Check whether the checks are independent, because dependent ones are worse than none.
  4. Reconciling one figure against an outside source establishes what no formula reading can.
  5. Ask the author which single cell, if wrong, would matter most.
  6. The honest output of a fast review is the list of parts nobody examined.
src

For further context, see Microsoft guidance on tracing formulas.