Syntha Lab RU

Descriptive statistics for retail: center, spread, outliers, intervals

A working set of statistical tools for sales and conversion: which average to use, how to measure spread, how to tell an outlier from an event, how to build a confidence interval and how many observations are needed so that a conclusion is not random. The decision it supports: whether and how far a number in a report can be trusted.

Math and analytics Brand and retail Verified: 2026-11-13

Open interactively →

Why this document

The analyst and the manager read this when sales or conversion figures must be judged as a real change or noise. The mean misleads when the data holds promotion spikes and disruptions, so you first choose the center (mean or median) and look at spread. An outlier may be a data error or a real event, and the action differs. The interval shows how far an estimate can be trusted, and the sample size tells how many observations to collect before concluding. Without this, one good sales week is taken for a trend.

At a glance

What is this document and why is it needed?
A working set of statistical tools for sales and conversion: which average to use, how to measure spread, how to tell an outlier from an event, how to build a confidence interval and how many observations are needed so that a conclusion is not random. The decision it supports: whether and how far a number in a report can be trusted.
Who owns the document and how often is it rebuilt?
Analyst (brand and retailer), as a method note behind reports: weekly in sales reviews, in every before/after comparison and when pilot stores are chosen; read by the buyer, the finance lead and the director.
What mistakes are the most common?
The mean of a series with promotion peaks is passed off as an “ordinary day” → inflated plan and staffing → show the median and mean together and tag events. Outliers are deleted silently → the total no longer matches the till → exclude only for the “norm”, compute totals on the full series. Store conversions are compared without intervals → noise is taken for an effect → compute the Wilson interval and the sample size in advance.

For whom

Brand
A brand works with series by partner, channel and own retail: seasonality and rare large orders make the mean unreliable, while the median and percentiles show the typical level. D2C and marketplace conversion is estimated with an interval, otherwise a difference of tenths of a percent is mistaken for an effect.
Retail
A retailer compares stores, days and weeks: a store’s daily sales are uneven across weekdays, and promotion peaks and disruptions (closures, repairs) stand out. Statistics separates the store norm from an event, compares store conversion fairly and sizes a pilot sample.

What it consists of

Center: mean, median, trimmed mean

  • The mean weights all days equally: in the example (56 days of one store) it is RUB 453 thousand against a median of RUB 415 thousand; the gap comes from two promotion days and one disruption.
  • The median splits the series in half and ignores single large values; it is shown next to the mean, not instead of it: total sales is a mean-based figure (total = mean × days).
  • The trimmed mean (dropping 10% at each end) and the mean without Tukey outliers give an intermediate estimate of an ordinary day; the cut rule is fixed in advance.

Spread: standard deviation, coefficient of variation, percentiles

  • The sample standard deviation (divisor n−1) is RUB 226 thousand; the coefficient of variation CV = SD ÷ mean = 49.8% shows spread relative to level and allows comparing stores and categories of different scale.
  • Percentiles describe the shape: p10 = 295, median = 415, p90 = RUB 604 thousand; “90% of days sell above p10” is readable without statistics and fits staffing and cash plans.
  • Brand: CV by SKU and partner feeds ABC-XYZ and safety stock (see the ABC-XYZ and safety stock documents); partner demand CV assesses order risk.
  • Retailer: CV by weekday is needed for shift schedules and the “day fell out of range” alarm; CV across stores of a group shows where the plan is unrealistic.

Outliers: rule and decision

  • Tukey rule: an outlier lies beyond Q1 − 1.5·IQR and Q3 + 1.5·IQR; in the example the fences are 87 and RUB 769 thousand, with 2 outliers (days 20 and 39). The rule only flags candidates; a person decides.
  • An outlier with a known cause (promotion, closure, POS failure) is tagged as an event and excluded from the “norm” but stays in the total; an outlier without a cause is first checked as a data error (duplicate receipts, wrong date).

Intervals and sample size

  • Confidence interval for the mean: mean ± t·SD/√n; for 56 days it is 393-RUB 514 thousand with outliers and 388-452 without them: outliers widened the interval 1.91 times.
  • Conversion: 1650 purchases from 18,400 visitors = 8.97%; the Wilson interval is 8.56-9.39%. A difference between two stores smaller than half the interval width proves nothing.
  • Sample size for conversion: n = z²·p(1−p)/e²; to pin down conversion of 9% within ±0.5 pp you need about 12,544 visitors, within ±1 pp about 3,136.

Brand and retailer: what is computed differently

  • Brand: observation unit: partner, channel, model or order; the series is short (weeks or months) and uneven, so the main tool is the median, percentiles and a bootstrap interval rather than the mean.
  • Retailer: observation unit: store and day; there are many long series, so t-interval estimates, like-weekday comparison and control by a group of similar stores are possible.
  • Brand: a large one-off partner order is an outlier by amount but not an error: it is tagged as an event and the norm is computed without it, while the purchase plan includes it as a separate line with a repeat probability.
  • Retailer: outliers are more often technical (POS failure, closure, duplicate receipts) or calendar-driven (promotion, holiday): data and the event calendar are checked first, then it is decided whether to exclude the day from the norm.
  • Brand: conversion is computed for D2C and marketplace channels separately; comparing with a partner without an identical visitor definition is invalid, so the definition (session, unique visitor) is written in policy.
  • Retailer: store conversion is receipts per counter visitors; counter accuracy (passers-by, staff) is estimated and added to the interval, otherwise the interval looks more precise than it is.
  • Brand: sample size determines how many partners or weeks are needed for a sell-out conclusion; with few partners a conclusion on the average is not possible and it is more honest to show each partner separately.
  • Retailer: sample size sets the number of stores and weeks for a pilot: comparing test and control on conversion needs thousands of visitors per group, so a pilot on two stores for a week proves nothing.
  • Brand: the brand owner is shown the median and a p10-p90 corridor by partners and channels, while the full CV table by SKU goes to the purchase planner.
  • Retailer: the retailer owner is shown the store norm (the weekday median) and a corridor, while the full statistics by day and percentiles go to the store director and the shift planner.

Where the data comes from

Data sourceFields
POS and traffic countersRevenue by receipt and day, receipt count, visitors, shifts; events (promotions, repairs, failures) in a separate register.
Brand sales recordsShipments and sell-out by partner, day, SKU; a “one-off large order” flag.
Event calendarPromotions, holidays, weather anomalies, closures; needed to explain outliers rather than delete them.

How to build it step by step

  1. Step 1. State the question and the unit of observation: “store sales per day”, “visitor conversion per week”, “order value”. Everything else depends on the unit.
  2. Brand: Step 2. Export the series by partner or SKU and move technical rows (returns) into a separate series.
  3. Retailer: Step 3. Export store sales by day excluding VAT with weekday and event flags.
  4. Step 4. Look at the histogram before any calculation: the shape (bell, right tail, two humps) decides which average and which intervals are appropriate.
  5. Step 5. Compute mean, median, SD, CV and percentiles p10, p25, p75, p90; a mean-median gap above 5-10% is a signal of skew or outliers (the threshold is illustrative and adjustable).
  6. Step 6. Find outlier candidates by the Tukey rule and match them with the event calendar; tag the cause or check the data.
  7. Step 7. If the series depends on weekday, compare like weekdays (Monday with Monday) rather than with the all-day average.
  8. Step 8. Build the confidence interval of the mean (t distribution) and of a share (Wilson interval); for the median and skewed series use the bootstrap (resampling with replacement, 1,000 repeats).
  9. Step 9. Before a pilot or comparison compute the required sample size and duration; if data is short, write “conclusion is preliminary” plainly.
  10. Step 10. Always show the number of observations, the center, the interval and the outlier rule in a report; the owner needs one line “norm and corridor”, the team needs the full table.

Formulas

MeasureFormulasNote
Mean and sample SDx̄ = Σx_i / n; s = √( Σ(x_i − x̄)² / (n − 1) )Divisor n−1 for a sample; n is acceptable for a full population (all days of the year).
Coefficient of variationCV, % = s / x̄ × 100Do not use when the mean is near zero or for series with negative values.
Percentile (linear interpolation)position h = (n − 1)·p; value = x_⌊h⌋ + (h − ⌊h⌋)·(x_⌈h⌉ − x_⌊h⌋) on the sorted seriesDifferent tools use different conventions; fix one across all reports.
Tukey fenceslower = Q1 − 1.5·IQR; upper = Q3 + 1.5·IQR; IQR = Q3 − Q1The 1.5 multiplier is conventional; for strongly skewed series use 3.
Interval of the meanx̄ ± t_(1−α/2, n−1) · s / √nRequires independent observations; days of one week depend on weekday, so work with weekly sums or with the same weekday.
Wilson interval for a sharep̃ = (p + z²/2n) / (1 + z²/n); half-width = z·√( p(1−p)/n + z²/4n² ) / (1 + z²/n)More reliable than the plain p ± z√(p(1−p)/n) for small n and small p.
Sample sizefor a share: n = z²·p(1−p) / e²; for a mean: n = (z·s / e)²For the mean of daily sales within ±5% with the s of the example about 382 days are needed; a guide for one independent group.

Formulas in notation

Mean and SD
x̄ = (1/n)·Σ x_i; s = √[ Σ (x_i − x̄)² / (n − 1) ]
Center and spread of a series; n−1 for a sample.
Outlier fences
L = Q_1 − 1,5·(Q_3 − Q_1); U = Q_3 + 1,5·(Q_3 − Q_1)
Points outside [L; U] are flagged as candidates.
Confidence interval of the mean
x̄ ± t_{1−α/2; n−1} · s / √n
Plausible range of the mean at confidence level 1−α.
Wilson interval
p̃ ± z·√[ p(1−p)/n + z²/(4n²) ] / (1 + z²/n), p̃ = (p + z²/(2n)) / (1 + z²/n)
Interval of a share (conversion) with good accuracy at small n.
Sample size for a share
n = z² · p(1−p) / e²
How many observations are needed for margin e.
Bootstrap of the median
CI = [ P_{2,5}(m*_1…m*_B); P_{97,5}(m*_1…m*_B) ]
m*_b: median of the b-th resample with replacement; B = 1,000.

Worked example

One store over 56 days: measures with and without outliers, RUB thousand

MeasureAll 56 daysWithout outliers
Days5654
Mean453.4420
Median415.5413.5
SD226116.3
CV, %49.827.7
Lower bound of the 95% interval of the mean392.9388.2
Upper bound of the 95% interval of the mean513.9451.7
Bootstrap interval of the median (lower / upper)371 / 460

The series is generated in code (seeded generator): a base day of RUB 420 thousand, weekday multipliers, noise and three artificial events. Outliers reduce precision and shift the mean by RUB 33.4 thousand; without them the interval of the mean is 1.91 times narrower. The data is illustrative, not a benchmark.

Figures

Histogram of daily store sales
Bar chart: 0–100 0, 100–200 1, 200–300 6, 300–400 16, 400–500 16, 500–600 11051015200–1000–100Days: 00100–200100–200Days: 11200–300200–300Days: 66300–400300–400Days: 1616400–500400–500Days: 1616500–600500–600Days: 1111600–700600–700Days: 44700–800700–800Days: 00800–900800–900Days: 00900–1000900–1000Days: 22days
Data as a table
MeasureDays
0–1000
100–2001
200–3006
300–40016
400–50016
500–60011
600–7004
700–8000
800–9000
900–10002

Histogram of daily store sales. Days per RUB 100 thousand bin: most days (49 of 56) lie between 200 and 600, two promotion days go into the right tail; day 28 with a disruption (RUB 113 thousand) is below the bulk but above the lower Tukey fence 87, so it is not an outlier. The right tail comes from days 20 and 39 (promotions, Tukey outliers); day 28 (disruption) sits in a bin left of the bulk.

Daily sales, mean and median
Line chart: Sales per day: 326 → 407; Mean (all days): 453 → 453; Median: 416 → 416; Mean without outliers: 420 → 42005001,0001,5002,0001611162126313641465156Sales per day, 1: 326Sales per day, 56: 407Mean (all days), 1: 453Mean (all days), 56: 453Median, 1: 416Median, 56: 416Mean without outliers, 1: 420Mean without outliers, 56: 420RUB thousand
  • Sales per day
  • Mean (all days)
  • Median
  • Mean without outliers
Data as a table
MeasureSales per dayMean (all days)MedianMean without outliers
1326453416420
2300453416420
3345453416420
4485453416420
5542453416420
6615453416420
7520453416420
8372453416420
9324453416420
10370453416420
11380453416420
12474453416420
13566453416420
14443453416420
15242453416420
16429453416420
17448453416420
18417453416420
19593453416420
201,805453416420
21511453416420
22224453416420
23344453416420
24366453416420
25353453416420
26614453416420
27532453416420
28113453416420
29327453416420
30354453416420
31299453416420
32446453416420
33504453416420
34561453416420
35403453416420
36342453416420
37414453416420
38317453416420
39907453416420
40452453416420
41666453416420
42502453416420
43263453416420
44276453416420
45343453416420
46432453416420
47595453416420
48445453416420
49486453416420
50410453416420
51324453416420
52291453416420
53413453416420
54519453416420
55643453416420
56407453416420

Daily sales, mean and median. The mean sits above the median because of outliers; the mean without outliers is closer to the median.

Conversion interval width and sample size
Line chart: Interval half-width, pp: 7.92 → 0.402468502501 0005 00020 000Interval half-width, pp, 50: 7.92Interval half-width, pp, 100: 5.6Interval half-width, pp, 250: 3.54Interval half-width, pp, 500: 2.5Interval half-width, pp, 1 000: 1.77Interval half-width, pp, 2 500: 1.12Interval half-width, pp, 5 000: 0.79Interval half-width, pp, 10 000: 0.56Interval half-width, pp, 20 000: 0.4pp
Data as a table
MeasureInterval half-width, pp
507.92
1005.6
2503.54
5002.5
1 0001.77
2 5001.12
5 0000.79
10 0000.56
20 0000.4

Conversion interval width and sample size. The half-width of the 95% interval falls as 1/√n: to halve it you need four times more visitors.

Visitors needed for a given conversion precision
Bar chart: ±0.25 50.2k, ±0.5 12.5k, ±1 3,136, ±2 784010k20k30k40k50k60k±0.25±0.25Visitors: 50.2k50.2k±0.5±0.5Visitors: 12.5k12.5k±1±1Visitors: 3,1363,136±2±2Visitors: 784784people
Data as a table
MeasureVisitors
±0.2550.2k
±0.512.5k
±13,136
±2784

Visitors needed for a given conversion precision. Margins of ±0.25, ±0.5, ±1 and ±2 pp at the example conversion.

Specifics by category

Apparel
Apparel sales depend strongly on weekday and weather: compute statistics within a season and like weekdays; at the season change the series breaks and a yearly mean describes nothing.
Footwear
A larger ticket and rarer purchases give a heavy right tail and few receipts per day; store intervals are wide, so draw conclusions by week and by store group.
Accessories
Gift peaks (February, March, December) create outliers that are the season norm; tag them as calendar events rather than deleting them, otherwise peak load is underestimated.
Kids
The school peak (late August-early September) and pre-holiday jumps produce a bimodal distribution; a single mean misleads, describe peaks and off-season separately.
Men
Demand is steadier, CV lower, outliers rarer: a median and an interval are enough; the repeat purchase share is higher, so conversion is computed separately for new and returning customers.
Women
Fashion waves and sales have a stronger effect: CV is higher, series are skewed, intervals are better built by bootstrap; compare full price and sale in separate series.

Common mistakes

  • The mean of a series with promotion peaks is passed off as an “ordinary day” → inflated plan and staffing → show the median and mean together and tag events.
  • Outliers are deleted silently → the total no longer matches the till → exclude only for the “norm”, compute totals on the full series.
  • Store conversions are compared without intervals → noise is taken for an effect → compute the Wilson interval and the sample size in advance.
  • Days of one week are treated as independent → the interval is too narrow → move to weekly sums or like weekdays.
  • CV is computed for a series with a near-zero mean or with returns → the number is meaningless → CV only for positive series, returns separately.

Russian specifics

  • For revenue statistics use amounts excluding VAT: the 22% rate from 2026 is one for most goods, but a rate change (and any 10% and 22% splits by group) distorts year comparisons; compare revenue excluding VAT. Rule: Tax Code Art. 164: Tax Code, Art. 164 (VAT rates).
  • Visitor data from cameras and Wi-Fi may contain personal data; for statistics store anonymized counters, not person identifiers. Rule: Law 152-FZ of 27.07.2006: Federal Law of 27.07.2006 No. 152-FZ on personal data (rule check not completed).

Open interactively →