AI in fashion analytics and content: what can be computed, limits, risks
The document covers where AI is genuinely useful in apparel, footwear and accessories (demand forecasts without history by attributes, recommendations, computer vision, content generation), how to check quality and which limits and risks to account for, including personal data under Federal Law 152-FZ. It decides whether to adopt a tool.
Why this document
AI is easy to show in a demo and hard to verify in daily work: a model can be confidently wrong, and content can infringe rights or expose customer data. The head of digital, the analyst and the lawyer read this document when they decide where to connect AI and where to keep a manual process. It sets quality checks on your own data and reminds about the personal data law 152-FZ. Without it a tool is adopted behind an attractive demonstration, and the problem is found when the mistake is already in buying or in a publication.
At a glance
- What is this document and why is it needed?
- The document covers where AI is genuinely useful in apparel, footwear and accessories (demand forecasts without history by attributes, recommendations, computer vision, content generation), how to check quality and which limits and risks to account for, including personal data under Federal Law 152-FZ. It decides whether to adopt a tool.
- Who owns the document and how often is it rebuilt?
- Brand: data director or analyst with the business process owner, pilots in a limited area and a quarterly review; the commercial director decides on launch with a lawyer. Retailer: analytics, IT and e-com; a separate person is responsible for personal data.
- What mistakes are the most common?
- The model is trained on sales instead of demand → stockouts and promotions are learned as the “norm” → clean or flag weeks without sizes and heavy promotions. No comparison with a simple baseline → a complex model is no better than the average → always compute the category mean as a control. The effect is measured in error percent rather than money → unclear whether accuracy is needed → convert the error into stock and shortage.
For whom
- Brand
- A brand applies AI to forecast new styles that have no history: attributes (type, price, color, material) are used to find similar items from past seasons and average their sales curve. Content (backgrounds, angle selection, descriptions) speeds up cards but needs control of color and rights. For a brand the key question is what is accepted as a decision and what remains a hypothesis.
- Retail
- A retailer uses AI in recommendations, search, size recognition and similar item matching, in traffic and staffing forecasts and in price monitoring through computer vision. Customer data (receipts, loyalty) falls under the personal data law, so design starts with minimization and anonymization.
What it consists of
What can really be computed
- Brand: the main task is forecasting a new model before production: batch size, MOQ, early and late order; an error here turns into stock at the brand or short supply to partners.
- Retailer: the main task is forecasting demand for a brand novelty for buying and allocation to stores; analogs are taken from the chain sales history, including other brands.
- A forecast for a new SKU by analogs: similarity by attributes, averaging the sales of k nearest, a check on past styles (backtest).
- Brand: recommendations work on the brand own site and app inside its range: sets, completing a look, repeat purchase.
- Retailer: recommendations may combine different brands into a look and use loyalty-card data; a basis for personal data processing and cannibalization control are needed.
- Recommendations: “bought together”, “similar items”, personal picks; the metric is revenue uplift and add-to-cart in A/B.
- Brand: vision tags the own catalog and campaign photos and checks sample-to-photo match; accuracy is checked on a sample for each category.
- Retailer: vision tags the catalog of third-party brands into a single attribute reference and supports image search and shelf compliance in store; tagging quality is checked against the brand card.
- Computer vision: attribute tags from photos, recognizing items on social media, display control, search by picture; accuracy is checked on a sample with manual labels.
- Brand: Content generation: backgrounds, looks, descriptions, translations; color and fit control and labeling where required are mandatory.
Quality assessment
- Forecast error: MAPE (mean absolute percentage error) and bias; comparison with a simple baseline (category average).
- Brand: value is counted in roubles of reduced stock of the own collection and shortage at partners; forecast accuracy is converted into production and top-up cost.
- Retailer: value is counted in roubles of reduced stock and lost sales by category and store, and reduced transfers between points.
- Value: accuracy gain in roubles (less stock and shortage), not in error percent.
- Control: a human in the loop for money decisions; a log of model and data versions.
Limits and risks
- Brand: data is cleaner in own channels, but wholesale shipments do not equal consumer demand; without partner stock data the model confuses shipments with demand.
- Retailer: data is richer by receipts and traffic, but size shortages, promos and transfers distort sales as demand; for third-party brands attributes are incomplete and need a reference.
- Data: few analogs, a heterogeneous catalog, promotions and stockouts distort “sales” as demand; without cleaning the model learns the error.
- Trends and novelty: a model of the past does not see a new trend; use it as a lower bound and complement it with the category manager decision.
- Law and ethics: personal data, copyright on training images and outputs, images of people, mandatory labeling of certain materials.
How to read the result and which decisions follow
- MAPE by analogs is 30%, by category average 38.9%; analogs are more accurate on 6 of 12 styles, bias -0.1%. Decision: on twelve styles the method does not prove superiority; a buying decision must not be given to a model, it is used as a second estimate alongside the buyer opinion.
- Brand: an analog forecast for a new style without history suits the first batch and the depth of the size run; after the first weeks of sales replace it with actuals. Decision: keep not only the forecast but the list of analogs so that misses can be reviewed.
- Retailer: analogs for you are your sales of similar items in your network; the brand sales data at other retailers are unavailable. Decision: first clean your data of weeks without sizes and heavy promotions, otherwise the model learns from shortage.
- Brand: images and descriptions for model training are protected by rights; use only materials with the right for such use and keep confirmations (part four of the Civil Code).
- Retailer: do not pass buyer personal data to external models without a legal basis and anonymization; for demand forecasting aggregates are sufficient.
Control checks
- Check: MAPE is computed over 12 styles, the error of each style equals |F - A| / A (code); the bias of -0.1% shows under-forecasting overall.
- Brand: the backtest on past styles must exclude the evaluated style from the analogs; otherwise the error is understated.
- Retailer: compare the model with a simple baseline (category average); if there is no gain, the complexity is not justified.
- Brand: MAPE is inflated on small actuals; show the absolute error in pieces and the error on the batch total next to it.
- Retailer: a human checks AI results in cards and descriptions before publication; responsibility for product information stays with you.
Where the data comes from
| Data source | Fields |
|---|---|
| Catalog with attributes | Type, price, color, material, silhouette, season, launch, role; a unified value dictionary. |
| Sales and stock by week | Sales, stock, size availability, price and discount by week; a stockout flag. |
| Images and descriptions | Photos with labels, usage rights, texts; source and date. |
How to build it step by step
- Step 1. Choose one task with measurable value (for example, a forecast for new styles), not “implement AI in general”.
- Retailer: Step 2. Bring the catalog to a single attribute dictionary; without it analogs cannot be found.
- Step 3. Clean the history: remove weeks without size availability and strong promotions or flag them; count demand, not sales.
- Step 4. Set a similarity rule and the number of analogs k; simple first, then complicate.
- Step 5. Run a backtest on past styles (each style against the rest): compute MAPE and bias and compare with the baseline.
- Step 6. Convert the error into money: how much stock and shortage a percent error produces; decide whether accuracy is enough for ordering.
- Step 7. Run a pilot in a limited area (one category, one season) with a human in the loop and control.
- Retailer: Step 8. For personal data set up grounds, purposes, retention and anonymization; record which data goes to a contractor.
- Brand: Step 9. For content introduce rules: color and fit check, image rights, labels for synthetic material, a ban on substituting the real item.
- Step 10. Review accuracy, bias and value quarterly; switch off a model that does not beat the simple baseline.
Formulas
| Measure | Formulas | Note |
|---|---|---|
| Brand: Analog forecast | Forecast = Mean sales of the k most similar items (over a comparable life period) | May be weighted by similarity; k is tuned on the backtest. |
| Brand: Similarity | Similarity = Σ weights of matching attributes (type, price, color, material) | Weights are set by an expert and checked by backtest; footwear adds the last, kids add the age group. |
| MAPE and bias | MAPE = mean(|Forecast − Actual| ÷ Actual) × 100; Bias = Σ(Forecast − Actual) ÷ ΣActual × 100 | MAPE is unstable at small actuals; also look at the revenue-weighted error. |
| Value of accuracy | Value = (Excess stock from baseline error − Excess stock from model error) × Cost + Lost margin from shortage | Counted per season; without this line accuracy stays academic. |
Formulas in notation
- Analog forecast
F = (1/k) × Σ_(j∈top k) s_j
The average sales of the k most similar items over a comparable period of life.- Similarity
sim(a, c) = Σ w_i × [attr_i(a) = attr_i(c)]
Weights of matched attributes: type, price, color, material (3, 2, 1, 1 in the example).- Mean absolute percentage error
MAPE = (1/n) × Σ |F_i − A_i| / A_i × 100
A_i is the actual; MAPE is sensitive to small actuals.- Bias
Bias = Σ (F_i − A_i) / Σ A_i × 100
Shows systematic over- or under-forecasting.
Worked example
Exercise: forecasting a “new” style from three analogs, tested on 12 past styles (conditional numbers)
| Style | Type | Actual, 8 weeks, units | Analog forecast | Error, % | Category mean | Error, % |
|---|---|---|---|---|---|---|
| No. 1 | Dress | 410 | 420 | 2.4 | 490 | 19.5 |
| No. 2 | Dress | 520 | 553 | 6.4 | 453 | 12.8 |
| No. 3 | Dress | 260 | 410 | 57.7 | 540 | 107.7 |
| No. 4 | Dress | 690 | 397 | 42.5 | 397 | 42.5 |
| No. 5 | Coat | 150 | 267 | 77.8 | 243 | 62.2 |
| No. 6 | Coat | 230 | 217 | 5.8 | 217 | 5.8 |
| No. 7 | Coat | 190 | 353 | 86 | 230 | 21.1 |
| No. 8 | Coat | 310 | 190 | 38.7 | 190 | 38.7 |
| No. 9 | Shirt | 480 | 567 | 18.1 | 530 | 10.4 |
| No. 10 | Shirt | 730 | 577 | 21 | 447 | 38.8 |
| No. 11 | Shirt | 560 | 577 | 3 | 503 | 10.1 |
| No. 12 | Shirt | 300 | 297 | 1.1 | 590 | 96.7 |
| Total MAPE | 30 | 38.9 |
Each of the 12 styles is forecast from the rest (leave-one-out). The analog forecast has a MAPE of 30% against 38.9% for the category mean; the bias of the analog forecast is -0.1%. There is a difference, but a sample of 12 styles is small: the conclusion “analogs are better” stays preliminary until checked on a bigger catalog and the next season. Numbers are conditional.
Figures
- By analogs
- By category average
Data as a table
| Measure | By analogs | By category average |
|---|---|---|
| No. 1 | 2.4 | 19.5 |
| No. 2 | 6.4 | 12.8 |
| No. 3 | 57.7 | 108 |
| No. 4 | 42.5 | 42.5 |
| No. 5 | 77.8 | 62.2 |
| No. 6 | 5.8 | 5.8 |
| No. 7 | 86 | 21.1 |
| No. 8 | 38.7 | 38.7 |
| No. 9 | 18.1 | 10.4 |
| No. 10 | 21 | 38.8 |
| No. 11 | 3 | 10.1 |
| No. 12 | 1.1 | 96.7 |
Forecast error by analogs and by category average. For each of 12 past styles: the error of the forecast by three analogs against the forecast by the category average.
- points
- regression line: y = 0 + 1·x
Data as a table
| Actual over 8 weeks, pcs | Analog forecast, pcs | |
|---|---|---|
| 410 | 420 | No. 1 |
| 520 | 553 | No. 2 |
| 260 | 410 | No. 3 |
| 690 | 397 | No. 4 |
| 150 | 267 | No. 5 |
| 230 | 217 | No. 6 |
| 190 | 353 | No. 7 |
| 310 | 190 | No. 8 |
| 480 | 567 | No. 9 |
| 730 | 577 | No. 10 |
| 560 | 577 | No. 11 |
| 300 | 297 | No. 12 |
Actual and analog forecast. Points on the diagonal mean an exact forecast; above it the forecast is too high, below too low.
- By analogs
- By category average
Data as a table
| Measure | By analogs | By category average |
|---|---|---|
| Dress | 27.3 | 45.6 |
| Coat | 52.1 | 31.9 |
| Shirt | 10.8 | 39 |
Mean error (MAPE) by item type. The analog method is not better than the category average everywhere: coats have the largest error.
Specifics by category
- Apparel
- Attributes for analogs: silhouette, fabric, color, season, price tier, role. Core has many analogs and the forecast is more accurate; fashion has few and the error is higher. So the backtest runs separately for core and fashion and is compared with a simple baseline (category average). Computer vision gives color and silhouette tags, but color in a photo is distorted by light: check on a sample. Generating backgrounds and looks is acceptable with color and fit control, replacing the real product is not.
- Footwear
- Attributes: type, last, heel height, upper and sole material, season. The forecast runs in two stages: model volume, then a split by the size curve; model volume is forecast more accurately than demand by each size. The main value is in size recommendation by foot length and last, because a size return is the most expensive. Generating footwear images is risky: the toe shape, sole and last get distorted and the customer sees a different product. EU, UK, US, Mondopoint labels are normalized to foot length in mm.
- Accessories
- There are almost no sizes, but material, shape, color, hardware and the gift occasion matter. The number of SKUs in a subcategory is small, so the forecast is noisy: a larger number of analogs k is taken or the forecast is made at subcategory level. “Bought together” recommendations work stronger than in apparel (bag and belt, scarf and hat). A season flag is mandatory because of gift peaks. For jewelry the image must not be generated or altered so that the impression of material and fineness changes.
- Kids
- Data about children is a heightened-risk area: personal data of minors requires consent of legal representatives, and profiling a child for advertising must be agreed with a lawyer. Size recommendations are built by height (6 cm step), not age. The forecast depends on the age cohort and the school calendar, the backtest runs by season: the school peak cannot be averaged with the rest of the year. Images of real children are not used for training and generation without consent; synthetic looks are labeled.
- Men
- There is less novelty and more repeat purchases of core, so the analog forecast is usually more accurate and the value of the model lies mostly in optimizing the size curve and the repeat order. Repeat-purchase recommendations (consumable basics: socks, T-shirts, underwear) work better than personal look selections. Size recommendations rely on fit (slim, regular) and garment measurements. Discount sensitivity is lower, so the price model assumes lower elasticity; verify on own data.
- Women
- More novelty and micro-trends, so the forecast error for fashion is higher and the value of a two-stage order and top-up is higher. Vanity sizing spread complicates size recommendation: brand size must be normalized to measurements, otherwise the model learns noise. Recommendations of looks and capsules give a bigger effect but depend more on catalog freshness. Computer vision helps find trends in social media; generated looks are labeled and color and fit are checked against the real product.
Common mistakes
- The model is trained on sales instead of demand → stockouts and promotions are learned as the “norm” → clean or flag weeks without sizes and heavy promotions.
- No comparison with a simple baseline → a complex model is no better than the average → always compute the category mean as a control.
- The effect is measured in error percent rather than money → unclear whether accuracy is needed → convert the error into stock and shortage.
- Personal data is passed to a contractor as is → a legal breach → anonymization, a contract, a data list.
- A generated look shows an item that is not the real one → returns and claims → a ban on substituting the item, check against the sample.
Russian specifics
- Processing customers personal data for recommendations and models requires a lawful basis, defined purposes, retention terms and protective measures (152-FZ); passing it to a contractor is arranged by a contract with a data list. Rule: Law 152-FZ of 27.07.2006: Federal Law of 27.07.2006 No. 152-FZ on personal data.
- Localization: recording, systematization, accumulation, storage, updating and retrieval of personal data of Russian citizens on collection are performed using databases located in Russia (Art. 18 part 5 of 152-FZ); separate requirements apply to cross-border transfer. The exact wording and exceptions are not checked against the law text (unverified): check with a lawyer. Rule: Law 152-FZ of 27.07.2006: Federal Law of 27.07.2006 No. 152-FZ on personal data (rule check not completed).
- Rights to images and training materials are intellectual rights (Part Four of the Civil Code): check sources, licenses and rights to the generated output; the legal status of an AI output should be clarified separately with a lawyer. Rule: Civil Code part 4: Civil Code, part four (Arts. 1232, 1235, 1237, 1486, 1491, 1515) (rule check not completed).
- Advertising materials created by AI fall under the same advertising law as ordinary ones: accuracy, labeling, identifier. Rule: Law 38-FZ of 13.03.2006: Federal Law of 13.03.2006 No. 38-FZ on advertising.