KPI dictionary and common definitions
One definition per measure: OTIF, forecast accuracy and its value, plan adherence; the formula, data source, frequency and owner.
About the topic
A metric is a promise that the same number means the same thing to everyone who looks at it. When purchasing counts on time by the date of shipment from the factory and logistics by the date of receipt at the warehouse, the meeting turns into an argument about whose number is right rather than a discussion of what to do. A metric dictionary cures exactly this: for every metric it fixes the formula, the base, the data source, the period, the owner and the decision that follows from the value.
Why it matters to a brand or retailer. Decisions on purchasing, budget, suppliers and plans rest on metrics, and if definitions drift, decisions drift too. Here you decide: which ten metrics truly steer the business, which definitions to fix in a policy, which thresholds to set from your own seasons and what to do when a metric enters the red zone.
The logic is built from the decision to the metric, not the other way round. Ask: which decision do I need to take, which number will support it, how should the number be computed so that it cannot be gamed, and who owns it. For example, to decide whether to change a supplier you need the share of orders on time and in full (OTIF), with a common definition of on time and in full. To decide whether to trust a forecast you need the forecast error and a comparison with a simple baseline forecast: if your forecast is no better than same as last week, it creates no value. To decide whether you have accumulated too much, you need purchase plan adherence.
The calculation base is mandatory: money metrics in Lab are counted excl. VAT, and next to the figure it is always stated whether it is with or without tax. The difference between 10 million roubles excl. VAT and 12.2 million with VAT at 22% is not an error but two different bases, and they cannot be compared.
Where to start: pick ten metrics on which you truly make decisions; for each write the formula, base, source, frequency and owner; compute OTIF, forecast error with forecast value and purchase adherence; set thresholds from past seasons and fix the dictionary so that only the owner can change it. Traps: calling OTIF different things in purchasing and logistics, comparing forecast accuracy without a naive baseline, changing a metric definition retroactively, mixing amounts with and without VAT, showing ten numbers without a threshold and an owner: no decision follows from them.
Key ideas
One definition per metric
Monthly revenue: the sales report gives 12.2 million roubles, the finance report 10 million. The difference is exactly 22%: in one case the price includes VAT at 22%, in the other it does not. If the base is not stated, a manager will pick the bigger figure while the budget will be built on the smaller. Agree the base and write it next to the number.
Six fields of a definition
Name, formula, base (with or without VAT, gross or net), data source and system, frequency, owner. Plus threshold and action: what we do when the value is red. Without an owner nobody will defend the definition, and without an action a metric turns into dashboard decoration.
On time: from which date?
100 orders: by factory shipment date 90 are on time, by warehouse receipt date 70, because 20 orders were delayed in transit. One metric, two conclusions: by the first the fault is logistics, by the second the factory is fine. Agree the date, the tolerance in days and the tolerance in pieces with the supplier in advance.
Forecast error and bias
Forecast for 4 weeks: 100, 120, 90, 110; actual: 110, 100, 95, 135. Absolute errors 10, 20, 5, 25, their sum 60 against actual total 440: weighted error (WAPE) 13.6%. Bias (forecast minus actual) −20 units, that is −4.5%: the forecast is slightly low. Error shows scatter, bias shows skew, and these are different problems.
Forecast value added: comparing with a naive baseline
A naive forecast same as last week for weeks 2–4: 110, 100, 95 against actual 100, 95, 135; errors 10, 5, 40, sum 55 against actual 330: 16.7%. Your model over the same weeks: errors 20, 5, 25, sum 50: 15.2%. Value added 16.7 − 15.2 = 1.5 points. That is modest: complicating the model is worth it only when the value is noticeable.
The change of a product is decomposed by factors with no remainder
Revenue grew by 1.69 m RUB: traffic −0.97 m RUB, conversion +2.63 m RUB, units per receipt −1.08 m RUB, price +1.10 m RUB.
Formula: Contribution = (Y₁ − Y₀) ÷ (ln Y₁ − ln Y₀) × ln(x₁ ÷ x₀)
What to do: Start with the factor of the largest modulus.
How to work through it
- List the decisions you take regularly and pick one metric for each.
- For each metric write the name, formula, base, source, frequency and owner.
- Agree contentious definitions (OTIF, revenue, stock, return) with those who compute them.
- Compute OTIF, forecast error with value added against a naive forecast and purchase plan adherence.
- Set thresholds from past seasons (not others' norms) and assign an action for the red zone.
- Fix the dictionary in a policy: only the owner may change a definition and only going forward.
- Reconcile metrics across sources (ledger, till, warehouse) and investigate gaps rather than average them.
Where to start
1) Choose ten measures you decide by. 2) Write the formula and the base for each (with or without VAT). 3) Name the data source and the system. 4) Set the frequency and the owner. 5) Calculate OTIF, WAPE with FVA and buying adherence. 6) Set your own thresholds from past seasons. 7) Fix the dictionary in a regulation and change it only by the owner’s decision. Calculations of the topic in order: “Forecast accuracy by week: window, bias and forecast value added”, “Driver decomposition of change (LMDI)”. See also (their primary topic is another one): “S&OP: forecast bias, error and the balance of demand and supply”, “Merchandise plan traffic light”, “Supplier OTIF and the lead-time stack”, “Forecast value added by stage and the price of error in roubles: the season assortment”.
Calculations in this topic
Forecast accuracy by week: window, bias and forecast value added
From weekly forecast and actual for a SKU or a group calculates the WAPE error and the bias overall and in a rolling window, the tracking signal, the comparison with a naive forecast (last week’s actual) and weeks with an error above your tolerance.
How accurate the forecast is by week, whether it is biased and whether it beats a simple repeat.
Driver decomposition of change (LMDI)
Decomposes the change between two periods of a metric that equals a product of factors (revenue = traffic × conversion × units per receipt × price) into the contributions of each factor. The logarithmic mean method gives contributions that add up to the change with no remainder and does not depend on the order of factors. It shows the contribution in money and in shares of the change and ranks the factors.
Which factor added or took how much from a metric between periods.
S&OP: forecast bias, error and the balance of demand and supply
For the monthly sales and operations planning cycle: reconciles the forecast with actual (bias, WAPE, tracking signal) and the supply plan with the demand plan. Shows whether the team overstates the forecast, where the error is large and when supply does not cover demand.
Does the team overstate the forecast, is the error large and is supply enough for demand.
Merchandise plan traffic light
Compares actuals with the plan on key merchandise-plan metrics and colors each green, amber or red by your thresholds. Shows what to look at first. The default thresholds are neutral (“an example”): there is no single Russian benchmark, set them from your own past seasons and targets.
Which plan metrics are already outside the tolerated deviation and where to look first.
Supplier OTIF and the lead-time stack
From supplier orders calculates the share on time, in full and both (OTIF), perfect orders, the worst supplier, average lead time planned and actual and by stage, and which stage used up the time buffer.
What share of orders arrives on time and in full and which stage uses up the time buffer.
Forecast value added by stage and the price of error in roubles: the season assortment
For groups of the assortment compares three forecasts of one season: naive (a repeat of the last season), a model (statistics or AI) and the result after manual adjustments. Counts the WAPE and bias of each stage, the value added of each step (FVA) and turns the error into roubles: over-forecast costs a markdown and holding, under-forecast costs lost margin. The definitions of WAPE, bias and FVA are given in the “Forecast accuracy by week” calculation (fcdash); here the point is the price of error and the stages.
Which stage of the forecast adds accuracy (the model or manual adjustments), how biased it is and how many roubles the error costs in the form of markdowns on the excess and lost margin.
Self-check questions
- What is the formula, base and source of each of the ten main metrics?
- Is each money figure with or without VAT and is it labeled next to the number?
- What share of supplier orders arrive on time and in full and who is worst?
- From which date is on time counted and how is it agreed with the supplier?
- Is your forecast better than a naive forecast and by how many points?
- Does the forecast have a skew in one direction and what does it mean for stock?
- By how much did purchasing deviate from plan and what does it mean for stock in days?
- What do we do when a metric is in the red zone and who decides?
- Did metric definitions change over the past year and is it recorded?
Knowledge check
Multiple-choice questions. The correct answer and the explanation follow each question.
What share does the OTIF measure count in Lab?
- A. Orders delivered on time ÷ all orders
- B. Orders delivered in full ÷ all orders
- C. Orders delivered on time and in full at once ÷ all orders
- D. Orders with no lateness, shortfalls, damage or document errors ÷ all orders
Correct answer: C. Orders delivered on time and in full at once ÷ all orders. OTIF = orders on time and in full ÷ all orders: the condition is “and”, so the measure is not above either the on-time share or the in-full share. The first two options are the separate “on time” and “in full” shares, the last one is the perfect order: it is stricter than OTIF because it also requires undamaged goods and correct documents.
Over a quarter a supplier handled 80 orders: 68 arrived on time, 72 in full, 62 on time and in full at once. What is the OTIF?
- A. 77.5%
- B. 85%
- C. 87.5%
- D. 75%
Correct answer: A. 77.5%. OTIF = 62 / 80 = 77.5%: the number of “both” orders is taken, it is entered as a fact. 85% is the on-time share (68 / 80), 87.5% comes from averaging the on-time (85%) and in-full (90%) shares, and 75% is the lower bound of “both” (68 + 72 − 80 = 60 orders): it can be estimated from the two shares but cannot replace the fact.
Actuals for weeks 1–5: 100, 120, 80, 100, 100 units; forecast: 100, 110, 90, 95, 105. The forecast is compared with the naive one (last week’s actual) on weeks 2–5. What is the forecast value added (FVA)?
- A. 14 pp
- B. 20 pp
- C. 7.5 pp
- D. 12.5 pp
Correct answer: D. 12.5 pp. The naive forecast on weeks 2–5 misses by 20 + 40 + 20 + 0 = 80 units out of 400: WAPE 20%. The forecast misses by 10 + 10 + 5 + 5 = 30 units: WAPE 7.5%. FVA = 20 − 7.5 = 12.5 pp: the forecast beats a simple repeat. 14 pp comes from taking the forecast WAPE over all five weeks (30 / 500 = 6%), while the comparison must be on the same weeks; 20 pp is the WAPE of the naive forecast alone, 7.5 pp is the WAPE of the forecast alone.
The forecast for 4 weeks: 520, 480, 500, 540 units; actuals: 500, 450, 520, 480. What is the forecast bias?
- A. −4.6%
- B. +4.6% (too high)
- C. 6.7%
- D. +4.4%
Correct answer: B. +4.6% (too high). Bias = (Σ forecast − Σ actual) ÷ Σ actual = (2,040 − 1,950) / 1,950 = +4.6%: in total the forecast is above the actual. 6.7% is the WAPE: (20 + 30 + 20 + 60) / 1,950, the errors are taken in absolute value and the sign is lost. +4.4% comes from dividing by the sum of forecasts (90 / 2,040): the base is always the actual. −4.6% is the opposite sign (actual minus forecast).
Buying counts OTIF by orders and gets 90%, logistics counts by order lines and gets 82%. The meeting argues whose number is right. What should be done under the rules of the KPI dictionary?
- A. Take the average, 86%, and show it to everyone
- B. Keep both figures under the name “OTIF”: each service looks at its own
- C. Fix one definition (the Lab base: orders) with a formula, a source and an owner, and keep the by-lines variant separately under its own name
- D. Take the stricter number, 82%, and replace the buying figure with it without a dictionary entry
Correct answer: C. Fix one definition (the Lab base: orders) with a formula, a source and an owner, and keep the by-lines variant separately under its own name. The dictionary fixes the formula, the base, the source, the frequency and the owner of a measure, and the definition can be changed only by the owner’s decision. The 86% average matches no definition; two figures under one name are the cause of the dispute; replacing with the stricter number without an entry repeats the same dispute in the next report. The by-lines variant stays useful, but under the name “OTIF by lines”.
To which metrics is this method applicable?
- A. To sums
- B. To shares of a sum
- C. To all without exception
- D. To products of factors
Correct answer: D. To products of factors. For sums a bridge is used. Covered in: The change of a product is decomposed by factors with no remainder.
What is the sum of contributions?
- A. Zero
- B. The change of the metric with no remainder
- C. The metric in the base period
- D. The average contribution
Correct answer: B. The change of the metric with no remainder. The logarithmic mean method. Covered in: The change of a product is decomposed by factors with no remainder.
Common mistakes
Mistake: Taking OTIF differently in purchasing and logistics
What it costs: Parties argue about numbers
How to avoid: One definition with a date and tolerances in days and pieces.
Mistake: Comparing forecast accuracy without a naive baseline
What it costs: A complex model looks good though it is no better than last week
How to avoid: Always compute forecast value added.
Mistake: Changing a definition retroactively
What it costs: The trend breaks and trust falls
How to avoid: Change going forward and record the date.
Mistake: Mixing metrics with and without VAT
What it costs: A 22% difference looks like success or failure
How to avoid: State the base next to the number.
Mistake: Showing ten numbers without a threshold and an owner
What it costs: No decisions follow
How to avoid: Each number gets a threshold, owner and action.
Entering zero factor values
What it looks like (illustrative figures): The logarithm is undefined.
Russia and the EAEU
There are no unified Russian standards for management metrics, so thresholds in Lab are not set: you choose them from your own past seasons. Count money metrics excl. VAT (the standard rate is 22% from 01.01.2026, law of 28.11.2025 No. 425-FZ, checked 14.11.2026; 10% for children’s goods on the list) and state the base next to the number. Companies on the simplified tax system have their own specifics: revenue for tax and management revenue may differ by recognition date, and in 2026 check the revenue threshold at which VAT arises with your accountant.
There are several data sources, and it is useful to reconcile them. Online-till receipts and withdrawal of labeled goods from circulation give an independent picture of sales; electronic documents show shipment and receipt dates and quantities; the warehouse system gives physical stock. A management report must explain differences with accounting, not hide them. Definitions of delivery time and penalties for failure should be fixed in the supply contract, not only in a policy, otherwise the metric has no force for the counterparty.
The dictionary principle does not depend on country; the base and currency change. Value added tax differs across union countries (reference as of 14.11.2026, check: VAT rates by country are on the page “EAEU and CIS countries”), so a metric revenue excl. VAT is comparable across countries and with VAT is not. To compare periods in different currencies decide in advance at which rate to convert (at the transaction date or the report date) and label it. For example, on a Kazakh platform fees have been shown excl. VAT since 05.01.2026 (reference, check); if your dictionary expects a fee including VAT, the margin metric will be counted differently. Add a country and currency field to the dictionary.
Categories and audiences
- Apparel
- In apparel forecast error is counted by model and colorway, not by SKU: at size level data is thin and noise is large. OTIF for fabrics and sewing is read separately because lateness causes differ. A single dictionary must know what sold at full price and maintained margin mean.
- Footwear
- In footwear the forecast must also be counted by size run: an error in popular sizes is more dangerous than in the extremes, and an overall WAPE does not see it. The return metric by size matters next to sales, so the definition of a return (by order date or by return date) is fixed in advance.
- Accessories
- For accessories demand is more often tied to gifts and promotions, so forecast error is large around peaks and small between them. Count it separately for peak and ordinary weeks, otherwise the average error misleads. The tolerance here is usually wider than for basic apparel.
- Kids
- Children’s goods are strictly seasonal: school and New Year. Forecast error and purchase adherence are read over those windows, and thresholds are set separately. The share sold by 1 September matters more than annual sell-through, because afterwards sizes lose value, and the metrics are counted by window and age group.
- Men
- For the men’s line show the forecast error separately for basics and novelties. For basics a narrow tolerance is appropriate, and comparison with a naive forecast “same as last year” is especially telling: if a complex model is no better, it is not needed. Novelties have no naive reference, they are judged against analogs, and bias is tracked separately.
- Women
- In the women’s line the dictionary must separate carried-over models from novelties: the error tolerance is wider for novelties and narrower for carried-over ones, otherwise one figure mixes the incomparable. Fix the definition of maintained margin separately (from which price, before or after returns) so that markdown and margin are counted the same way in all reports.
Questions this topic answers
- What are the formula, base and source of each of the ten main measures?
- What share of supplier orders is on time and in full and who is the worst?
- Is our forecast better than simply repeating last week?
- How far has buying deviated from plan and what does it mean for stock?
- What do we do when a measure is in the red and who decides?
Key measures and terms
Study first
What next
Next: “Data model of standard reports”, “The planning calendar”, “Stock accuracy, shrink and availability”.
Documents on this topic
- S&OP: cycle and documents
- Management report pack
- Organization and incentives: RACI, KPI/OKR, bonuses, store payroll
- Descriptive statistics for retail: center, spread, outliers, intervals
- Cohort analysis: retention table, curve, LTV forecast and RFM as mathematics
- Price and volume indices: Laspeyres, Paasche, Fisher, chain indices, deflator and the 4-5-4 calendar
- Systems and data: 1C and ERP, PIM and MDM, BI, integrations (e-document exchange, marking, marketplace API), cost of ownership and data quality
- Forecast accuracy and the value of an AI forecast: WAPE, bias and FVA on a season assortment
- Warehouse and data: capacity for the peak, stock and sales reconciliation, returns, RFID and slotting