Stock accuracy, shrink and availability
Whether the records match the shelf, what shrink and mismatches cost and how much revenue is lost to missing goods.
About the topic
All stock formulas, from the reorder point to the purchase plan, silently assume one thing: the number in the system equals what is on the shelf. When it does not, the error does not stay on a single line. The system does not order an item that has “three left” on paper while the shelf is empty, and it orders one that is in fact well stocked. The topic is about how to measure the gap, what it costs, and how to tell loss of goods from unavailability of goods.
There are three different but connected questions. Record accuracy: the share of items in the books that match reality in a sample count. Shrink: how much goods disappeared through theft, damage and receiving or shipping errors, and how much of gross profit it eats. Availability: how often the item could be bought and how many sales went because it was not there. They form a chain: poor record accuracy leads to wrong orders, wrong orders to stock-outs, stock-outs to lost revenue.
The reasoning rests on a sample, not on a full check. Take 200–500 lines, count them on the shelf, compare with the system and get the share that matched. For example, 316 of 400 lines matched: accuracy is 79%, so a fifth of the items are “lying” to the system. Shrink is measured in one unit, either at cost or at retail, otherwise stores cannot be compared. With turnover of RUB 120 million, gross margin of 52% (RUB 62.4 million) and shrink of RUB 1.5 million at cost, shrink eats about 2.4% of gross profit. Lost revenue is counted from what the item sold on the days it was in stock: if a fast seller makes RUB 50,000 a day, 8% of days without stock in a month is RUB 120,000, and after allowing for the 40% of shoppers who took a substitute, RUB 72,000 of revenue remain, about RUB 37,000 of gross profit.
Start by counting a small sample of key items in one store and the warehouse. Typical traps: checking only the warehouse total without sizes and colors, counting shrink at retail in some places and at cost in others, computing sales pace over days without stock (the result is zero), and carrying one store’s result over to the whole chain.
Key ideas
Record accuracy
Accuracy by lines and accuracy by money can diverge. Of 500 lines, 410 matched, which is 82%. But if the mismatched lines carry RUB 3.2 million of a RUB 10 million stock, accuracy by value is 68%. That is why value accuracy is counted next to line accuracy, and the expensive class A items are counted more often.
Phantom stock
The system shows 3 units, the shelf has 0. While the record is not fixed, replenishment does not start and the item may not sell for months. That is why fast sellers with zero sales and positive stock are checked first.
Shrink: one unit of measure
Store A lost RUB 0.9 million at cost on a turnover of RUB 60 million: 1.5%. Store B lost RUB 1.2 million at retail price on a turnover of RUB 100 million: 1.2%. B looks better. But if retail is twice the cost, A’s loss at retail is RUB 1.8 million, which is 3.0%, and A is worse than B not by 0.3 but by 1.8 percentage points.
Availability and lost revenue
Availability is the share of time when an item can be bought. Lost revenue is counted from the sales pace on days in stock, minus substitution: some shoppers will take another color or size. Without subtracting substitution the loss is overstated.
Cost of a discrepancy
A discrepancy costs money not by itself but through the decisions built on it: an extra order freezes cash, a missed order creates a shortage. So the price is counted in roubles through order errors and lost sales, not just in units.
An improvement over the naive baseline is not yet a gain
An AI forecast is compared with the rule “as in the last period”, not with zero. Over 11 periods the AI error is 12.1%, the naive one 14.4%: an improvement of 15.8% at a threshold of 10%, the AI is more accurate in 82% of periods. But the gain in money is RUB 20,400 (34 units less error × RUB 600) against the tool cost of RUB 20,000: the result is RUB 400. The improvement threshold is passed, but the payback is on the edge. The definitions of error, bias and forecast value added are given in the terms “wape” and “pd-fva” and in the “Forecast accuracy by week” calculation (fcdash); what matters here is the economics of the replacement.
Formula: Improvement = 1 − AI error ÷ naive error; gain = saved error × price of error − tool cost
What to do: Introduce an AI forecast as a one-season pilot while the result after the tool cost is close to zero.
Think first: Improvement 15.8% at a threshold of 10%. Introduce it without conditions?
Answer: No: after the tool cost RUB 400 remains. A pilot and a recount at the price of error are needed.
Actual sales understate demand while the shelf is empty
The accuracy of the forecast and of stock are linked: if goods ran out, sales are below demand and the under-forecast error looks smaller. Suppose 1,320 units were sold, the goods were in stock 80% of the days: the demand estimate is 1,320 ÷ 0.8 = 1,650 units. For a forecast of 1,200 units the error by the actual is 120, by demand 450.
Formula: Demand estimate = sales ÷ the share of days in stock
What to do: To assess the forecast take demand by days in stock and not bare sales: otherwise a shortage looks like accuracy.
How to work through it
- Define the scope of the check: one store and the warehouse, key categories, one week.
- Take a random sample of 200–500 lines at the size and color level, not only at the model level.
- Calculate the share of matched lines and split errors into “books above reality” and “books below reality”.
- Estimate the cost of discrepancies: excess orders, missed orders, phantom stock on fast sellers.
- Compare stores by shrink on a single base and find where it exceeds the threshold: in one store or in all.
- For key items calculate availability and lost revenue with a substitution adjustment.
- For wholesale add the share of order lines fulfilled: it is availability from the customer’s point of view.
- Assign actions: counting frequency by ABC class, fixing causes, a repeat check a month later.
Where to start
Run a sample count of 200–500 lines and enter the result in “Stock record accuracy”; then “Shrink” by store and “Availability” for key items. Calculations of the topic in order: “Stock record accuracy and the cost of mismatches”, “Stock shrink: share, threshold and cost to profit”, “Availability and lost revenue”. See also (their primary topic is another one): “Defects: losses, supplier recovery and a reserve”, “Footwear returns: reasons, reverse logistics, the share back on sale and the loss”, “Safety stock by fill rate and lead-time variability”, “Checking an AI forecast against a naive baseline: WAPE, bias, improvement, share of periods and the gain net of costs”.
Calculations in this topic
Stock record accuracy and the cost of mismatches
From a sample count it calculates record accuracy by lines and by value, the net balance of mismatches and estimates the lost revenue from “phantom” stock: an item is on record but not on the shelf.
How far stock records match reality and what the mismatches cost.
Stock shrink: share, threshold and cost to profit
From book and counted stock by store it calculates shrink, its share of revenue, the comparison with your threshold and what part of gross profit it eats. Shows the worst store and a breakdown by cause.
How much goods we lose, whether it is above our threshold and what part of gross profit it eats.
Availability and lost revenue
Calculates the share of days an item could be bought, estimates lost sales, revenue and margin by item (adjusted for substitution) and, for wholesale, the shares of order lines and pieces filled.
How often the item could be bought and how much revenue and margin we lost without it.
Defects: losses, supplier recovery and a reserve
By category counts the defect rate, the loss at cost, the proceeds from selling defective items (second grade), the supplier’s recovery, the net loss and an estimate of a reserve for the next period. Shows where defects cost most.
What defects cost, how much the supplier will return and what reserve to set aside.
Footwear returns: reasons, reverse logistics, the share back on sale and the loss
Splits footwear returns by reason (size, defect, did not like), calculates reverse logistics and handling, what share went back on sale as new and at a reduced price (worn returns), how much is lost on the price reduction and write-offs and what one return costs.
What footwear returns cost, what part goes back on sale as new and worn, and what cutting size returns would give.
Safety stock by fill rate and lead-time variability
Counts safety stock in two ways: by cycle service level (the probability of not running out) and by fill rate (what part of demand is met from the shelf), with demand and lead-time variability; shows reorder points and the money in stock.
How much safety stock to hold to meet the required share of demand and how that differs from protecting every order cycle.
Checking an AI forecast against a naive baseline: WAPE, bias, improvement, share of periods and the gain net of costs
From a series of actual sales and an AI forecast for the same past periods it builds a naive forecast (the value of the period t − lag: last week or the same period of the last season) and compares with it. It counts the WAPE and the bias of the AI forecast and of the naive one, the improvement against the baseline, the share of periods where the AI is more accurate, whether the improvement threshold is passed, the saved error in units and the gain in money net of the cost of the tool. It answers whether the AI forecast deserves more trust than a simple rule. The forecast must have been made before the fact and not fitted; the default data are illustrative.
Whether the AI forecast is more accurate than the simple rule “as in the last period”, by how much, in how many periods and whether it pays off after the cost of the tool. WAPE and bias are counted by the definitions of the “Forecast accuracy by week” calculation (fcdash).
Self-check questions
- Do the books match the shelf at size and color level, not only at model level?
- What is the share of matched lines and in which direction do the errors go?
- Are there fast sellers with zero sales at a positive stock record?
- Is shrink in different stores counted on one base, at cost or at retail?
- What share of gross profit does shrink eat and where is it above the threshold?
- Are days with zero stock excluded when calculating sales pace?
- Is substitution accounted for: some shoppers take another color or size?
- Has the sample been checked in several stores, not only one?
- Does the stock in the books match the marking system for marked goods?
Knowledge check
Multiple-choice questions. The correct answer and the explanation follow each question.
The AI error is 12.1%, the naive baseline error 14.4%. What is the improvement over the baseline?
- A. 2.3%
- B. 84.2%
- C. 12.1%
- D. 15.8%
Correct answer: D. 15.8%. 1 − 12.11 ÷ 14.38 = 15.8%. The difference in points (2.3%) and the ratio (84.2%) are not the improvement. Covered in: An improvement over the naive baseline is not yet a gain.
The AI saved 34 units of error, the price of error is RUB 600, the tool RUB 20,000. What is the result after costs?
- A. 20,400 ₽
- B. 400 ₽
- C. 20,000 ₽
- D. 40,400 ₽
Correct answer: B. 400 ₽. 34 × RUB 600 = RUB 20,400; minus RUB 20,000 = RUB 400. The gain without subtracting the tool (RUB 20,400) overstates the result. Covered in: An improvement over the naive baseline is not yet a gain.
The improvement threshold is passed, the result after costs is RUB 400. What is sensible?
- A. introduce without review
- B. refuse: the threshold is not passed
- C. a one-season pilot with a recount of the price of error
- D. add one more tool
Correct answer: C. a one-season pilot with a recount of the price of error. There is an improvement, but the payback is on the edge: an error in the price of error or in the periods flips the sign of the result. A pilot is cheaper than a rollback. Covered in: An improvement over the naive baseline is not yet a gain.
1,320 units were sold, the goods were in stock 80% of the days. What is the demand estimate?
- A. 1,056 units
- B. 1,320 units
- C. 1,650 units
- D. 1,584 units
Correct answer: C. 1,650 units. 1,320 ÷ 0.8 = 1,650. 1,056 results from multiplying instead of dividing; 1,584 is simply plus 20%. Covered in: Actual sales understate demand while the shelf is empty.
The forecast is 1,200 units, the demand by days in stock is 1,650 units. What is the forecast error by demand?
- A. 120 units
- B. 1,650 units
- C. 450 units
- D. 1,200 units
Correct answer: C. 450 units. |1,200 − 1,650| = 450. 120 results from comparing with bare sales on an empty shelf. Covered in: Actual sales understate demand while the shelf is empty.
600 units were sold, the goods were in stock 50% of the days. What is the demand estimate and what must be remembered?
- A. 1,200 units; the estimate is rough: it assumes that demand on the days without goods is the same
- B. 300 units
- C. 600 units, availability does not matter
- D. 900 units
Correct answer: A. 1,200 units; the estimate is rough: it assumes that demand on the days without goods is the same. 600 ÷ 0.5 = 1,200. At such availability the estimate rests on the assumption of equal demand across days, so it is marked as approximate. Covered in: Actual sales understate demand while the shelf is empty.
Common mistakes
Mistake: Checking only the warehouse total
What it costs: Discrepancies by size and color cancel out and stay invisible
How to avoid: Count at size and color level.
Mistake: Counting shrink at retail in some stores and at cost in others
What it costs: Stores cannot be compared
How to avoid: Fix a single base.
Mistake: Counting sales pace over all days, including days without stock
What it costs: Pace is understated and so is the order
How to avoid: Count over days with stock.
Mistake: Carrying one store’s count over to the whole chain
What it costs: Shrink and accuracy depend on staff, layout and traffic
How to avoid: Check a sample across several stores.
Mistake: Not subtracting substitution in lost revenue
What it costs: The loss is overstated and justifies excess stock
How to avoid: Subtract the share of shoppers who took another variant.
Mistake: The AI forecast is compared with itself
What it costs: Nice accuracy without a baseline Tool cost 20,000 ₽ against a gain of 20,400 ₽.
How to avoid: Compare with the naive rule and count the result after costs Set the naive baseline, the improvement threshold and the price of error before the pilot; look at the result after costs.
What it looks like (illustrative figures): An AI error of 12.1% looks good, but the naive rule gives 14.4%; after the tool cost the result is RUB 400.
Mistake: Forecast accuracy is counted by sales with an empty shelf
What it costs: Under-forecast is understated and the model seems better The error is understated by 3.8 times and the shortage stays unnoticed.
How to avoid: Estimate demand by days in stock. Record the days without goods and recalculate sales into demand before comparing with the forecast.
What it looks like (illustrative figures): Sold 1,320, in stock 80% of the days, forecast 1,200: an error of 120 against 450 by demand.
Russia and the EAEU
Stocktaking in Russian accounting follows accounting rules: agree the procedure, the papers and the booking of surpluses and shortages with your accountant. Shortage within the norms of natural loss is written off to expenses, beyond the norms it falls on the responsible persons or other expenses depending on the grounds; check the procedure.
For marked goods (footwear, apparel, light industry), on top of the discrepancy in the books there is a discrepancy between your stock and the stock in the Chestny ZNAK system: it is a problem in itself, since goods without a code in circulation cannot be sold. At receiving, check codes against the universal transfer document (UPD) in electronic document exchange, and reflect customer returns and write-offs in the marking system. A separate point for online sales: on marketplaces the stock is seen by the marketplace’s warehouses, and a gap between your books and the marketplace warehouse produces cancellations and penalties under the platform’s rules; check the terms in your contract. A weekly sample count of key items and a full stocktake once a year or quarter are usually cheaper than a full weekly check.
The principles of record accuracy and shrink are the same in all countries. Accounting and e-document exchange differ: Belarus and Kazakhstan have national electronic invoices (a guideline, verify), Russia uses the UPD and exchange through operators. Marking of stock follows national rules: in Kazakhstan mandatory marking starts on 01.12.2026 with stages through 2027, and stock is to be marked by 01.01.2028 (a guideline, verify); for Belarus, Armenia and Kyrgyzstan the dates: indicative as of 14.11.2026, check the text of the act, verify. The stocktaking procedure and the write-off of shortages are set by national accounting and tax law: ask a local accountant.
Categories and audiences
- Apparel
- In apparel errors pile up at the size and color level: the model total matches while the core size is empty. A huge number of SKUs and frequent try-ons leave sizes mixed on hangers. Count the core sizes of fast models more often and at size and color level.
- Footwear
- In footwear a box can be empty or hold a different pair, and the left and right shoe can differ in size. Losses come from display samples and fitting. Lost revenue is higher: a size is not replaced by a neighbor. Check that the pair is in the box and that the size on the label matches.
- Accessories
- Accessories are small, expensive per gram and sit near the till: theft losses are above average, especially for jewelry, gloves and eyewear. Records often have no sizes, and errors happen at model and color level. The sample is smaller, counts are more frequent, and anti-theft measures (tags, display cases) pay back faster.
- Kids
- Children’s goods have many small SKUs, sets and seasonal sizes, and more returns: a parent buys a size “to grow into” and exchanges. Receiving and return errors build up toward the 1 September season. For children’s products safety and marking are checked more strictly, so a gap with the marking system is especially unpleasant.
- Men
- In the men’s basics a typical source of record error is mix-up: one T-shirt in three sizes and two colors, a box put into the wrong cell. On a fast size a zero quietly turns into a lost sale. Count accuracy by size, not by model, and start with the core of the run.
- Women
- The women’s range has more SKUs, silhouettes, colors and short collections: records are harder, discrepancies higher, and confusing color or size is easier. There are many try-ons and returns, and part of the goods sits in the hall as “tails”. Count accuracy by size and color and run a count after sales, when the run is especially mixed.
Questions this topic answers
- Do the records match what is on the shelf?
- How much money do we lose to mismatches and shrink?
- Where is shrink above the norm: in one store or in all?
- How many sales do we miss through missing goods?
Key measures and terms
Study first
What next
Next step: “Omnichannel: click and collect and ship from store”. Related: “Replenishment and stock cover”, “Order data quality”, “Capital, multi-brand and the planning cycle”.
Documents on this topic
- Merchandise stock balance
- Warehouse: WMS, handling cost, SLA, returns and cycle counting
- Loss prevention: shrinkage, counting, fraud
- Footwear returns and warranty: reasons, calculation, procedure, defects by component
- Fashion tech: digital product passport, RFID, AR try-on, 3D design
- Fit and sizing: anthropometry, ease, size runs, returns
- Demand forecasting: smoothing, seasonal indices, analogs and error control
- Holt-Winters, seasonal decomposition and Croston: forecasting with trend, season and intermittent demand
- Returns, non-pickup and fines on platforms: causes, losses by size, defects and logistics
- AI in fashion and analytics: where to apply it, where not, how to check the result, which data are needed and which errors occur. A memo and a checklist
- Forecast accuracy and the value of an AI forecast: WAPE, bias and FVA on a season assortment
- AI in fashion analytics and content: what can be computed, limits, risks