Stock Audit Sampling Methods: How Auditors Select Items Without Counting Everything

How an Auditor Decides What to Count

An auditor who cannot count every line counts in three layers: every line that carries serious value, a sample drawn from the rest, and a handful of items picked out because something about them looks wrong. Each layer answers a different question, and a sound report keeps them apart.

The framework comes from SA 530, the auditing standard on audit sampling issued by ICAI. It permits both statistical and non-statistical sampling; the auditor designs the sample around the assessed risk and selects items so that every unit has a chance of selection. Misstatements found in the sample are then projected to the population they were drawn from, rather than treated as the only errors that exist. A lender's stock audit is a special-purpose engagement, so the standard is applied there as the professional benchmark rather than by mandate, but the logic is the same.

The starting point is not the stock list but the question the report has to answer. A conclusion about the value of stock offered as security needs selection weighted towards value. A conclusion about whether the stores function records movements correctly needs selection spread across locations and handling points. Most of the defects in sampling come from using one kind of sample to support the other kind of conclusion.

1. Count the top of the list in full. Lines that each carry a large share of the value are examined completely, because sampling them saves almost no time and leaves a material amount unverified if one is missed.
2. Sample the remainder in bands. The rest of the population is grouped by value or location and items are drawn from each band, so no band goes unexamined.
3. Add items that invite attention. Lines that have not moved, rates that look wrong or stock held somewhere unusual are added on judgement and reported as findings in their own right.

What the sample can support is a statement about the population within stated limits. What it cannot do is prove that every line not selected is correct, and a report that implies otherwise is claiming more than its evidence reaches.

Value-Weighted and ABC Selection

Stock populations are almost always skewed: a small share of the lines carries most of the money. Value-weighted selection is built around that fact, so the counting effort follows the value instead of the line count.

Ranking comes first. Each line's quantity is multiplied by its rate, the list is sorted from highest to lowest value, and a running total shows where the value sits. Grouping the ranked list into A, B and C bands is the common way of formalising it: the A band holds the few lines that carry the bulk of the value, the B band the middle, and the C band a long tail that adds up to comparatively little.

An example with round figures shows why this works. Suppose a register of 1,200 lines is worth Rs 4 crore and the top 90 lines are worth Rs 3 crore between them. Counting those 90 lines in full verifies three quarters of the value from under a tenth of the lines, and a sample from the remaining 1,110 lines then tests the rest. The same register counted line by line from the top of the location list could reach a third of the lines and still miss most of the value.

Random and Systematic Selection

Value bands decide how much attention each part of the stock receives. Within a band, something else has to decide which items are examined, and that is where random and systematic selection come in.

The bias that randomness protects against is rarely dishonesty. An auditor left to choose freely drifts towards stock that is accessible, well labelled and easy to count, and those are the items least likely to be wrong. A random draw reaches the top shelf, the back of the godown and the mixed stack as readily as the front row.

• Random selection draws items using random numbers against the band's list, so every line has a known chance of being chosen. It is the easiest method to defend and the one statistical conclusions depend on.
  Suits a conclusion about the band
• Systematic selection takes every nth line from a sorted list after a random start. It is quick and easy to document, but if the list has a repeating pattern the interval can line up with it: a list sorted by location, with the same number of lines in each rack, can produce a sample drawn entirely from one shelf level.
  Check the list order first
• Haphazard selection means picking without a conscious pattern. It is acceptable for a small non-statistical sample, but it is not suitable where the conclusion is to be stated statistically, because nobody can show afterwards that the choice was free of bias.
  Small samples only

Combining a value band with a random or systematic draw inside it is what makes the selection defensible. The band sets the coverage, the draw sets the items, and the working papers record which method was used in which band so that the selection can be re-performed by a reviewer.

Judgemental Selection and Its Limits

Alongside the structured sample sits a layer of items chosen because something about them invites a closer look. This layer often produces the most useful findings, and it is also the layer most often misreported.

Typical reasons for adding an item on judgement:

• a line that has shown no movement for many months, which may be obsolete or may not exist;
• the same description repeated across several lines, which can hide duplicates or a split balance;
• stock recorded at a location the team was not expecting, including a job worker's premises or a rented godown;
• a rate that is out of line with recent purchase invoices for the same item;
• anything the previous audit flagged and the response did not fully resolve.

A judgemental pick supports a finding about the item examined. If that line is wrong, it is wrong, and the finding stands on its own. It supports nothing about the population, because the items were chosen precisely because they were not typical. Strictly speaking it is not a sample at all.

Problems start when the two layers are merged into one coverage figure. A report that adds judgemental picks to the structured sample and quotes a single percentage overstates the basis for its conclusion on the balance. Keeping them apart costs one line in the method section and lets the reader weigh each layer for what it is.

Coverage the Report Has to State

Coverage is a disclosure, not a detail. A report that leaves it out is asking to be relied on further than its evidence goes.

Banks generally accept sample-based verification, with the coverage they expect set in the stock audit scope they issue, often a high share of value. That makes the coverage statement the first thing a credit officer reads after the headline figure, and it should answer five questions plainly:

1. Value covered. The share of total stock value actually verified, against the figure in the stock statement being tested.
2. Lines covered. The number and share of lines examined, stated beside the value figure rather than instead of it. In a skewed population the two differ widely, and quoting only one flatters the work.
3. How items were chosen. The method in each band, the value above which lines were counted in full, and which items were added on judgement.
4. What was left unverified. Whether the remainder is many small lines, which is ordinary, or a few large ones that could not be reached, which is a limitation the reader has to know about.
5. What the differences mean for the whole. Errors found in the sample projected to the band they came from, and what was done when the projected figure was too large to accept.

That last point is where samples earn their keep. When the differences found in a band, projected across the band, exceed what the reader can tolerate, the answer is not to report the sample's errors and move on. The sample is extended, the band is counted in full, or the report says clearly that the band's figure is unreliable. A report stating high value coverage while a material line went unverified has answered the arithmetic and not the question.

Setting a Sample You Can Defend

A sample is defensible when its basis was fixed before anyone counted, it matches the conclusion the report draws, and a reviewer could repeat the selection from the working papers.

• Match the sample to the conclusion. A statement about total value needs selection weighted towards value. A statement about whether the counting and recording process works needs selection spread across locations, categories and handling points, including the cheap lines. Decide which conclusion is wanted before deciding what to count.
  Decide the question first
• Record the basis before counting. The method, the full-count threshold, the bands and the reasons for judgemental additions are written down in advance. A basis described afterwards cannot be told apart from one chosen to fit the result.
  Written before the count
• Know when to count everything. A small population where sampling saves little, a balance already in dispute where any sample will be challenged, and a band where one missed line would be material are all cases where a full count is the only answer that holds.
  Full count

For a borrower, the practical point is that the selection basis is part of what the lender relies on. If your bank's scope asks for a stated coverage, or a previous report was questioned for its sampling, agreeing the basis before the team travels avoids a dispute afterwards. That is settled when a stock audit is scoped, together with the sites, the stock classes and the report format the bank wants.