How to calculate attribute elasticity to optimise assortment mix

Merchandisers have long known that a navy jumper sells differently from a rust one, that cashmere commands a premium over acrylic, and that a £45 price point captures a different customer than £75. What they lack is a systematic method to quantify those differences and use them to guide buy plans. Attribute elasticity, the statistical measure of how individual product features affect demand, offers that method. By running regression analysis on historical sales data, companies can isolate the impact of colour, fabric, silhouette, price and other attributes, then apply those coefficients to forecast sales for new styles and optimise the mix of features in an assortment.

The technique matters now because fashion cycles have shortened and the share of newness in every season has risen. A large share of every season's assortment has no sales history at all, making standard forecasting methods struggle with fashion's specific complexity. In a McKinsey survey of merchants published in January 2026, 71 per cent said AI merchandising tools have had limited to no effect on their business so far. The gap sits in the foundations: sales history quality, attribute consistency, and financial reconciliation. Fixing those foundations unlocks the value that regression based attribute models promise.

Half of executives in the BoF McKinsey State of Fashion 2025 Executive Survey plan to localise their go to market model and value proposition, especially through pricing, fulfilment channels and assortment, and 65 per cent plan to alter their assortments to include a variety of options across price points. Both require understanding how specific features perform in specific channels and geographies, which is precisely what attribute elasticity reveals.

What is attribute elasticity and why does it differ from price elasticity?

Attribute elasticity measures the percentage change in demand resulting from the presence or change of a specific product feature, holding other features constant. A colour elasticity of 0.15 for olive green, for example, means that olive green products sell 15 per cent more units than the baseline colour, all else equal. Price elasticity, by contrast, measures how demand responds to a change in price. Both are elasticities because they express sensitivity, but they apply to different levers.

Fashion retailers care about attribute elasticity because most merchandising decisions involve choosing features, not just setting prices. A buyer selecting which five colours to offer in a new style needs to know which colours drive volume. Fast fashion requires precision at the style colour size channel level, each combination carrying a distinct demand signal with its own lifecycle. Attribute elasticity quantifies that signal before the first unit ships.

Price elasticity of demand is used to capture the phenomenon where, when the price drops, the quantity demanded of the product increases. Determination of price elasticity of demand is a challenging task in fashion e commerce, as there is a wide range of products that are sold, so retailers experience all types of elasticity. Attribute elasticity adds another dimension: it separates the effect of the product's intrinsic features from the effect of its price.

How do you structure sales data to run a regression on product attributes?

The unit of analysis is the SKU or SKU-week pair, depending on whether you are modelling total season sales or weekly sell through. Each row in the dataset represents one product, and the columns contain the dependent variable, sales units or revenue, and the independent variables, which are the attributes. Attributes must be coded consistently. Colour names need a standard taxonomy: is burgundy a red or a purple? Fabric composition should aggregate to meaningful categories such as natural, synthetic, or blend rather than listing every fibre permutation separately.

The factors influencing demand include Color Group, NOS (Never Out of Stock, indicating basic products intended to be continuously available), Fit, Collar Type, Thickness, and a product group is defined by a unique combination of attributes including Year, Week, Color Group, NOS, Fit, CollarType, and Thickness. In addition to visual embeddings, structured product attributes such as brand, price, fabric type, and seasonality are incorporated into the representation space, and these attributes are concatenated with the visual features to construct a multimodal embedding.

Price is typically included as a continuous variable, though some analysts bin prices into tiers. Categorical attributes such as colour, fit, or neckline are encoded as dummy variables, with one category omitted as the baseline. If a style comes in five colours, you include four dummy variables; the omitted colour becomes the reference against which the others are measured. Missing data is a practical problem: if fabric weight was not captured for 30 per cent of SKUs, those rows either need to be excluded or the missing values need to be flagged with a separate dummy variable.

External variables can improve the model. The dataset provides a comprehensive set of features including product image, category, dominant color, fabric type, seasonality, launch date, brand, and retail price, and it incorporates time series data such as sales trends and stock availability post launch, and to enhance the forecasting models with contextual demand signals, the dataset also includes auxiliary information like Google Trends data for categories, colors, and fabrics. Weather data, promotional activity, and competitor pricing can all be added as control variables if the data exist.

Which regression method should you use and why does it matter?

Ordinary least squares (OLS) linear regression is the starting point for most attribute analysis. It is transparent, fast to run, and produces coefficients that are easy to interpret. The linear regression fit achieves a strong R squared of 0.6 to 0.8 and the learned coefficients exhibit intuitive semantic structure, for example premium brands and specialized commodity categories contribute positively to price, while generic or high volume subcategories contribute negatively, and this decomposition generalizes well to unseen data, indicating that unit prices in grocery retail indeed follow an additive structure over product metadata, which is further supported by interpretable SHAP analyses and bar plots of regression coefficients.

Linear regression assumes that the effect of each attribute adds to the baseline demand. A product with both a desirable colour and a desirable fabric gets the sum of both effects. This additive assumption is often reasonable, but not always. If certain attributes interact, for example if floral prints only drive demand in lightweight fabrics, not heavy ones, you need interaction terms. An interaction term is the product of two dummy variables, and its coefficient captures the additional effect when both attributes are present together.

When the dependent variable is sales units and many SKUs sell zero units in a given week, a Poisson or negative binomial regression may fit better than OLS because it accounts for the count nature of the data and the excess zeros. Logistic regression is appropriate if the goal is to predict whether a SKU will sell above a threshold rather than to predict the exact sales level. Ridge or lasso regression can help if you have many attributes and some are correlated, as both methods shrink coefficients and lasso can drop irrelevant features entirely.

For assortment optimisation, the multinomial logit (MNL) model is widely used. The MNL model involves customers associating a preference weight with each product, and the preference weight can increase via synergy, which occurs between a pair of products when the retailer offers both of them in his assortment even when customers purchase at most one product, and the purchase probabilities of either product may increase or decrease depending on the increase in preference weight and the weights of the other product. The MNL framework translates attribute coefficients into choice probabilities and expected revenue, making it possible to optimise which products to offer in a constrained assortment.

How do you interpret the coefficients and apply them to new products?

A regression coefficient tells you the change in the dependent variable associated with a one unit change in the independent variable, holding all other variables constant. For a dummy variable such as colour equals olive, the coefficient is the difference in sales between olive products and the baseline colour. If the baseline is black and the coefficient for olive is minus 200 units, olive products sell 200 fewer units on average than black products with otherwise identical attributes.

Each coefficient represents the effect of that predictor while holding all other predictors constant, and product quality has the largest coefficient, meaning a one unit improvement in perceived product quality is associated with a point increase in satisfaction, assuming support speed, price fairness, and website ease stay the same. For continuous variables such as price, the coefficient represents the marginal effect. If the price coefficient is minus 50, each additional pound of price is associated with 50 fewer units sold.

To forecast demand for a new style, multiply each of its attribute values by the corresponding coefficient and sum them. A new jumper in olive, cashmere, crew neck, priced at £65 would get the baseline intercept, plus the olive coefficient, plus the cashmere coefficient, plus the crew neck coefficient, plus £65 times the price coefficient. The result is the predicted sales volume. This is analogous item modelling. A new style's forecast borrows the sales curve of similar past styles, matched on attributes like silhouette, fabric weight, and price point, and this is the standard answer for fashion items, and its accuracy equals the quality of the similarity judgment.

Confidence intervals around each coefficient indicate statistical significance. If a coefficient's confidence interval includes zero, the attribute may not have a reliable effect and should either be dropped or the model should be re estimated with more data. Standard errors tend to be large when sample sizes are small or when an attribute appears in only a few SKUs.

What does Zara's attribute driven approach reveal about practical implementation?

Trend information flows daily into a database at head office, which is used by designers to create new lines and modify existing ones, Zara warehouses have standardised product information with common definitions allowing quick and accurate preparation of designs with clear manufacturing instructions, and all its designers are given unparalleled independence in approving products and campaigns based on daily data feeds indicating which styles are popular. This is attribute analysis in action: every design decision is informed by how similar attributes performed.

Store staff are trained to observe and report, and if multiple customers in a Mumbai or Milan store ask for a different collar, a longer hem, or a blue version of an existing top, that information reaches designers at headquarters within days, and the store is the research lab. Zara's Point of Sale systems track what sells and what doesn't at a granular level down to individual store and neighbourhood, inventory logs identify slow movers before they become a dead stock problem, and if a product is performing in Seoul but not in Madrid adjustments happen inside the current cycle not the next season.

Viral designs can hit shelves in as little as 17 days, this blend of creativity and analytics keeps sell through rates above 85 per cent, far higher than the industry average, and by only producing what sells Zara avoids markdowns protects margins and ensures customers always see something they want. Inditex delivered robust H1 2026 results with sales up 7.6 per cent, or 9.2 per cent in constant currency, to EUR 19.8 billion and net income rising 6.8 per cent to EUR 3 billion. In the first half of 2026 gross profit increased 8.3 per cent to reach EUR 11.6 billion and the gross margin reached 58.7 per cent.

The Zara model demonstrates three principles that generalise beyond fast fashion. First, attribute data must be captured at the moment of sale, not reconstructed later. RFID tags and modern point of sale systems make this feasible. RFID tags let Zara track every single garment in real time from factory to distribution centre to store shelf to fitting room, and when you pick up a jacket and try it on that signal feeds into inventory data. Second, the attribute taxonomy must be consistent across design, buying, and merchandising. If the design team calls a colour sage and the buying team calls it mint, the regression will treat them as separate and dilute both coefficients. Third, the feedback loop must be fast enough to inform decisions that still matter. Weekly updates are useful; monthly updates often arrive too late.

How should price be included in the model and what are the limits of price elasticity?

Price can be included as a linear term, a logged term, or as price tiers, depending on the distribution and the theoretical relationship you expect. A linear term assumes each additional pound has the same effect on demand, which is rarely true. Logging price allows diminishing marginal effects: the difference between £30 and £40 matters more than the difference between £130 and £140. Price tiers, such as under £50, £50 to £100, and over £100, let the model estimate a separate elasticity for each segment.

Data from Q1 2026 shows average price increases slowed to 2.1 per cent globally down from 9.4 per cent in 2022 to 2023, while brands like Gucci and Burberry adjusted entry level pricing downward by up to 8 per cent to re engage aspirational buyers. The defining feature of the 2026 luxury reset is a strategic pivot away from blanket price inflation toward segmentation, as between 2020 and 2023 brands relied heavily on price increases to offset supply chain disruptions and capitalize on post pandemic demand, however by late 2024 demand elasticity began to show cracks especially among middle tier luxury consumers in Europe and China.

Price elasticity varies by category and customer segment. Demand for ethically sourced products is highly elastic and sensitive to price with sustained willingness to pay primarily for higher priced luxury items, which suggests that aside from luxury items demand for ethically sourced products may be highly price sensitive. Mass market apparel is typically elastic: a 10 per cent price increase may reduce volume by 15 per cent. Luxury goods are often inelastic or even Veblen goods, where higher prices signal status and can increase demand. Giffen Veblen goods are a type of product for which the demand increases when there is an increase in the price, and in the fashion environment retailers not only experience different types of elasticity but also have to deal with Giffen Veblen goods, for example a Rolex watch is a type of Veblen good as its demand increases with an increase in price because it is used more as a status symbol.

Including price in an attribute regression lets you control for price effects when estimating the impact of other features. Without controlling for price, a cashmere coefficient might be inflated because cashmere items are also more expensive, and higher priced items often target a different customer. The regression separates the cashmere effect from the price effect. However, price and certain attributes may be collinear: if all floral prints are in the lowest price tier, the model cannot reliably separate the floral effect from the low price effect.

What data quality and taxonomy issues will break the model?

Attribute elasticity models fail when the input data are inconsistent, incomplete, or poorly defined. The most common problem is taxonomic inconsistency: the same colour called navy by one buyer, midnight by another, and dark blue by a third. The gap sits in the foundations: sales history quality, attribute consistency, and financial reconciliation, and the algorithms are rarely the weak link. Unless the data are cleaned and standardised, the regression will treat these as three separate colours, fragment the sample, and produce unreliable coefficients.

Missing data is another trap. If fabric composition was not recorded for older seasons, those SKUs either need to be excluded, reducing sample size, or the missingness needs to be modelled. If the missing data are not random, for example if fabric was only recorded for higher priced items, excluding them biases the sample. A missing data dummy variable can sometimes help, though it assumes that the missing category is meaningful.

Sample size matters. If only 12 SKUs out of 800 are yellow, the yellow coefficient will have a large standard error and may not be statistically significant even if yellow truly drives demand. Rare attributes need either more seasons of data or should be aggregated into broader categories. Overfitting is a risk when the model includes more attributes than the data can support: the model fits the noise in the training sample but fails to generalise.

The requirement is data not scale, and retailers need clean sales history by store and a consistent attribute taxonomy. Establishing that taxonomy is not a technical task but a governance task. It requires cross functional agreement on how to define and name every attribute, and it requires discipline to enforce those definitions during product creation. Many companies discover that their product information management system allows free text entry, which defeats standardisation.

How do you move from coefficients to actionable assortment decisions?

Once you have reliable coefficients, the next step is to use them to simulate alternative assortments and select the one that maximises revenue, margin, or sell through subject to constraints such as budget, minimum breadth, or supplier capacity. The input to the assortment model is the forecasted sales, item attributes, selling price and cost price of the article at a store level, and the objective is to obtain the list and quantity of articles that is to be kept in the store for the particular phase such that overall revenue is maximized and the business constraints are satisfied.

A simple approach is to rank styles by predicted sales or predicted margin, then select the top N that fit within the budget. This greedy heuristic works well if products do not cannibalise each other, but it ignores substitution effects. If two olive jumpers are both in the top 20, offering both may not double the volume because they compete for the same customer. The multinomial logit model accounts for substitution by modelling each customer's choice among the offered assortment.

Integer programming formulations can handle more complex constraints. The constraints include the maximum distinct articles in a store, total budget at sub class level, and minimum and maximum distinct articles at sub class level. The decision variables are binary: offer this SKU or not. The objective function is the sum of forecasted sales times margin for all selected SKUs. Modern solvers can handle thousands of SKUs and dozens of constraints, though run time increases with problem size.

Localisation is important. If a sleeve style or color resonates in Madrid but not in Stockholm Zara increases production where demand exists rather than shipping the same assortment everywhere. Running separate regressions by region or store cluster lets you estimate region specific coefficients and tailor assortments accordingly. This requires sufficient regional sales data, which is often available for large chains but not for brands with limited distribution.

What are the risks, limits and counter arguments to attribute driven forecasting?

Attribute elasticity models assume that past patterns will continue, which is fragile in fashion where trends shift rapidly. A colour that drove demand last autumn may be passé this autumn. The model cannot predict a trend break unless it includes forward looking signals such as social media sentiment or search volume. Historical averages carry almost no predictive weight in a category where a silhouette can peak and fade within a single selling month.

Correlation is not causation. A high coefficient for a particular neckline might reflect an unobserved factor such as the fact that all those styles were featured in marketing campaigns or worn by influencers. If you forecast demand for a new style with that neckline but without the marketing support, the model will overestimate. Controlling for promotional activity in the regression helps, but unmeasured confounders remain a risk.

Attribute models are reductionist. They quantify what can be measured but ignore what cannot, such as brand heat, cultural moment, or the intangible appeal of a particular design. Pre season planning still needs human judgment for trend direction, brand positioning, and category strategy, and AI handles the volume of work: attribute based forecasting, cluster recalculation, replenishment, and markdown optimization. A design that feels fresh or tells a story may outperform its predicted coefficient by a wide margin.

Sample size limits generalisation. If the training data come from a single channel or a single customer demographic, the model's predictions may not hold in a different context. A colour that performs well in stores may underperform online if the screen rendering differs from the physical garment. Regional differences in taste mean that a coefficient estimated on UK data may not apply in Japan. Running separate models by channel, region, or customer segment addresses this but fragments the data and reduces statistical power.

Over optimisation is a risk. If every competitor uses attribute models and everyone converges on the same colours and fabrics, the assortment becomes homogeneous and customers have less reason to choose one brand over another. Differentiation requires taking calculated risks on attributes that the model does not favour. Customers are expected to become more cautious with their spending in 2026, and for those who do splurge they will be paying special attention to signifiers of value for money such as craftsmanship durability and sharp creative direction, and brands that raise prices without improving quality or design risk alienating consumers and eroding brand equity.

Frequently asked questions

What should fashion companies do now?

1. Audit your attribute taxonomy across design, buying, merchandising and IT systems. Identify where the same attribute is named differently or where definitions are ambiguous, then establish a single controlled vocabulary and enforce it in your product information management system.
2. Clean and consolidate historical sales data at the SKU level with attributes appended. Remove SKUs with missing or inconsistent attribute data unless you have a method to impute or flag them. Aim for at least two full seasons of data covering a representative range of attributes.
3. Run a pilot regression on one category where you have the most confidence in data quality, such as knitwear or denim. Use OLS to start, include price and the most important attributes, and check the R squared and coefficient significance. If results are poor, revisit data quality before adding complexity.
4. Validate predictions against holdout data by training the model on past seasons and testing it on the most recent season. Compare predicted sales to actual sales for each SKU and calculate the mean absolute percentage error. If accuracy is below 70 per cent, investigate which attributes or segments are driving the error.
5. Integrate attribute forecasts into the buying process by providing buyers with predicted sales for each candidate style based on its attributes. Treat the forecast as decision support, not a mandate, and collect feedback on where the model was wrong so you can refine it.
6. Build cross functional governance to maintain data quality over time. Assign ownership of the attribute taxonomy to a single team, establish a change control process for adding new attributes, and run quarterly audits to catch drift or inconsistency.
7. Invest in capability, not just software by training merchandisers and planners to understand regression output, interpret coefficients, and challenge forecasts when they conflict with market intelligence. Technical tools deliver value only when the people using them understand their assumptions and limits.