Quantitative Assortment Optimization

Introduction and context

Assortment planning determines which stock-keeping units (SKUs) a retailer lists, how much space they receive, and when they should be retained, replaced, or removed. In a competitive retail environment, the central problem is not simply to offer more products. It is to construct an assortment that satisfies customer missions while producing attractive sales, margin, inventory productivity, and space productivity.

Quantitative assortment optimization converts this problem into a structured decision process. It combines SKU productivity, Pareto analysis, demand transfer, substitutability, incrementality, space elasticity, operational constraints, and scenario analysis. The resulting decision is managerial rather than purely mathematical: models provide evidence, but category strategy, customer needs, supplier relationships, and risk must also be considered.

The analysis is especially important in categories with limited shelf capacity, high SKU proliferation, short product life cycles, or significant substitution. Delisting a weak SKU may release space and inventory investment, but it may also transfer demand to competitors, reduce choice, or damage the category’s price image. Conversely, retaining every SKU can create duplication, poor availability, excessive working capital, and confusing choice.

Learning objectives

By the end of this chapter, a learner should be able to:

1. Calculate and interpret SKU productivity and related performance measures.
2. Use Pareto analysis without confusing concentration with causation.
3. Distinguish demand transfer, substitution, and incremental demand.
4. Estimate and interpret space elasticity.
5. Formulate assortment decisions subject to commercial and operational constraints.
6. Compare listing, delisting, and rationalization scenarios.
7. Critically evaluate data, model risk, governance, and ethical implications.

1. The assortment optimization problem

A category assortment has a finite number of product positions, shelf facings, display locations, inventory capacity, and buyer attention. Each SKU contributes several types of value:

• Economic value: sales, gross margin, contribution, and cash generation.
• Shopper value: need fulfilment, choice, convenience, and perceived quality.
• Strategic value: traffic generation, differentiation, price-image support, and brand relationships.
• Operational value: replenishment reliability, handling simplicity, and supply-chain compatibility.

A SKU with low direct sales may still be strategically important. For example, a premium item may establish a price ladder, while a specialist product may fulfil an important dietary or cultural need. Therefore, optimization should not maximize one metric mechanically.

A useful conceptual objective is:

Here, R is category revenue, G is contribution margin, C is customer or mission coverage, S is strategic value, I is inventory investment, and K represents operating or complexity cost. The coefficients α, β, γ, δ, λ, and μ are managerial weights. This expression is a decision framework rather than a universal formula. It is useful only when the measures are defined consistently and normalized where necessary; revenue, percentage coverage, and monetary cost cannot be added meaningfully without suitable scaling.

2. Data foundation and unit of analysis

Before applying a model, the analyst must define the observation period, store cluster, channel, and performance measures. A national average may conceal substantial differences between urban convenience stores, hypermarkets, and online fulfilment centres.

Minimum data requirements usually include:

• weekly unit and value sales by SKU and store;
• regular and promotional prices;
• gross margin and supplier funding;
• stock-on-hand, stock-outs, and lost-sales estimates;
• shelf facings, space, and display location;
• product attributes and brand architecture;
• store, season, and customer-segment identifiers.

Sales observed during stock-outs are censored demand: low sales may reflect unavailable inventory rather than weak customer preference. Promotion can also inflate apparent productivity. Analysts should therefore distinguish baseline sales from promotional sales and use comparable periods where possible.

A practical unit of analysis is SKU × store cluster × week. This permits estimation of local variation and prevents a high-volume store from dominating all conclusions. Products should also be grouped into comparable roles, because comparing a staple, a seasonal gift, and a niche specialist solely on unit sales is invalid.

3. Measuring SKU productivity

3.1 Core measures

The most common measures are:

Gross margin is normally calculated as sales minus cost of goods sold. Contribution may additionally subtract variable fulfilment, handling, or markdown costs. The denominator and time period must be stated. For example, annual sales per facing should not be compared with weekly sales per facing.

For a SKU i, sales productivity can be written as:

where S_i is sales value during the selected period and F_i is average shelf-facing count. The measure assumes that facings are measured consistently and that the period includes comparable demand conditions. It does not prove that adding facings will increase sales; it is descriptive, not automatically causal.

3.2 Worked example: productivity ranking

Suppose three SKUs produce the following quarterly results:

Sales per facing are 10,000 for A, 12,000 for B, and 12,000 for C. SKU A is the largest revenue contributor, but B and C are more space productive. If the retailer is space constrained, B and C deserve investigation. However, C may have lower absolute contribution and higher replenishment complexity. The decision cannot be made from one ratio alone.

4. Pareto analysis and the long tail

Pareto analysis ranks SKUs by a selected measure, calculates cumulative contribution, and identifies concentration. If SKUs are ranked by descending sales, cumulative sales for the first k products are:

Here, S_i is sales of SKU i and n is the total number of SKUs. The often-used “80/20 rule” is a heuristic, not a law. A category may have 70/30, 90/10, or a relatively even distribution.

Pareto analysis is useful for identifying core products, reviewing tail productivity, and setting differentiated service levels. It becomes misleading when managers assume that every SKU below a cumulative threshold should be delisted. A tail item may provide incremental demand, protect a customer segment, complete a range architecture, or reduce the risk of customers switching retailers.

The analysis should be repeated for units, revenue, margin, transactions, and customer missions. A product that is weak in revenue may be strong in margin or attract a valuable segment. Managers should also conduct a sensitivity test: rankings based on one abnormal month or a major promotion may be unstable.

5. Demand transfer, substitutability, and incrementality

5.1 Why delisting effects are difficult to predict

When a SKU is removed, its previous sales do not necessarily disappear. Some demand transfers to other products in the category, some transfers to another retailer, some is postponed, and some represents purchases that would never have occurred without the SKU.

Let baseline sales of delisted SKU d be S_d. Let t_j be the proportion transferred to retained SKU j, and let L be the proportion lost from the category or retailer. Then:

The residual, 1 − L − ∑t_j, may represent postponed demand, unobserved channels, or estimation error. If transfer is expressed in units, transferred units to SKU j equal t_j S_d. The transfer rates must be estimated from comparable delistings, shopper research, substitution matrices, or controlled tests.

A delisting is financially attractive only if the retained sales and cost effects compensate for lost contribution. A simplified incremental contribution calculation is:

where m_j is contribution per unit of retained SKU j, m_d is contribution per unit of the delisted SKU, A is avoided cost or released value, and ∆K is any additional cost caused by the change. The equation assumes stable prices and margins; if price, promotion, or availability changes, those effects must be modelled separately.

5.2 Substitutability

Substitutability is the degree to which customers regard two products as alternatives. It depends on brand, pack size, flavour, quality, price, dietary attributes, format, and shopping mission. A product may be physically similar but commercially non-substitutable.

A substitution matrix can classify relationships as high, medium, or low. A more advanced approach uses cross-price elasticity. For products i and j:

A positive cross-price elasticity often indicates substitution, while a negative value may indicate complementarity. In retail data, interpretation requires caution because prices are endogenous: retailers may discount products precisely when demand is weak, and promotions often affect several SKUs simultaneously.

5.3 Incrementality

Incremental sales are sales that would not have occurred without the SKU or activity. Total category sales are not equal to the sum of incremental contributions if products cannibalize one another. A premium variant may shift customers from a standard variant rather than expand the category.

Incrementality can be estimated through:

• matched-store or geographic tests;
• before-and-after analysis with a control group;
• customer-level basket analysis;
• structural demand models;
• surveys combined with observed behaviour.

A robust test should control for seasonality, price, promotion, distribution, stock-outs, competitor activity, and store trends. Correlation between a new SKU and category growth is not sufficient evidence of incrementality.

6. Space elasticity and shelf allocation

Space elasticity measures the responsiveness of sales to a change in allocated space. A commonly used local approximation is:

If sales rise from 1,000 to 1,100 units while facings rise from 2 to 3, the percentage changes are 10% and 50%, giving ε_s = 0.20. This indicates diminishing response in that observed range, but it does not imply that every additional facing will produce the same result.

A log-log model is often used when sufficient data are available:

Here, Q_i is sales, F_i is space or facings, X_i is a vector of control variables such as price, promotion, store size, and season, and u_i is the error term. The coefficient ε_s is interpreted as an elasticity when the model is appropriately specified. The model assumes a stable relationship over the observed range and requires attention to endogeneity: high-selling products are often assigned more space, so simple regression may overstate the effect of space.

Space allocation should also respect physical realities. Minimum facings may be required for visibility; maximum facings may be limited by pack dimensions; and shelf capacity may be integer-valued. A mathematically optimal fractional facing is not operationally feasible.

7. Optimization constraints and decision models

A practical assortment model may use binary decision variables. Let x_i = 1 if SKU i is listed and 0 otherwise. A basic listing constraint is:

where N is the maximum number of listings. If f_i is the number of facings assigned to SKU i and F is total capacity:

Typical constraints include:

• minimum number of brands or pack sizes;
• required representation of dietary, cultural, or regulatory segments;
• minimum order quantities and case-pack requirements;
• supplier or private-label commitments;
• minimum availability targets;
• budget and inventory limits;
• warehouse and replenishment capacity;
• integer facings and planogram adjacency;
• legal restrictions on regulated products.

A stylized mixed-integer model is:

subject to capacity, budget, range, and service constraints. Here, v_i represents estimated net value of listing SKU i, and w_i represents the marginal value of space. These values must already account, as far as possible, for cannibalization and demand transfer. Otherwise, the model double-counts sales.

Optimization is not a substitute for judgement. Hard constraints should be distinguished from soft constraints. For example, legal compliance is hard; a preferred brand balance may be soft. Constraint relaxation can reveal whether a recommendation is driven by a genuine commercial trade-off or by an arbitrary parameter.

8. Scenario analysis

Scenario analysis tests how decisions perform under alternative assumptions. At minimum, an assortment review should compare:

1. Current assortment: the baseline.
2. Productivity-led rationalization: remove low-productivity SKUs.
3. Customer-coverage scenario: retain or add products serving underrepresented missions.
4. Space reallocation scenario: retain listings but alter facings.
5. Risk scenario: assume lower transfer and higher lost demand.

A scenario table should report revenue, contribution, availability, inventory, SKU count, customer coverage, and operational complexity. For example:

Indices are useful for comparison but do not replace absolute values. A balanced scenario may be preferred because it protects coverage while improving contribution, even if it is not the revenue maximum.

Sensitivity analysis varies uncertain parameters such as transfer rates, space elasticity, margin, and stock-out correction. If a SKU is retained only under an optimistic transfer assumption, the decision is fragile. Expected-value analysis can be used when probabilities are defensible:

where p_s is the probability of scenario s and V_s is its value. Probabilities should not be invented merely to create precision; ranges and break-even analysis may be more honest.

9. Illustrative case: FreshBite ready meals

Background

FreshBite, an illustrative supermarket chain, sells 60 chilled ready-meal SKUs in 120 urban stores. Shelf capacity is fixed at 48 facings per store. Management believes that the tail is excessive, but customer complaints indicate that vegetarian and lower-calorie options may be underrepresented.

Problem

The lowest 12 SKUs account for 4% of category sales and 3% of contribution. A mechanical delisting would release 12 positions, but managers are uncertain about substitution and customer loss.

Analysis

The team cleans 26 weeks of data, removes weeks with severe stock-outs, and separates promotional sales. It groups products into standard, premium, vegetarian, vegan, low-calorie, and family-size roles. A matched-store test delists four comparable low-volume SKUs in 20 test stores while retaining them in 20 control stores.

Results suggest that 72% of delisted unit demand transfers to retained products, 18% is lost from FreshBite, and 10% is postponed or unexplained. Transfer is highest within the same dietary and price segment. The test also shows that one apparently weak vegan SKU has high mission coverage and attracts baskets containing other high-margin products.

Approach and solution

FreshBite does not delist all 12 tail products. It removes eight products with low contribution, high duplication, and low transfer-adjusted value. It retains the vegan SKU and three products needed to maintain price and dietary architecture. Six released facings are allocated to high-velocity products, four to under-spaced vegetarian products, and two remain as flexible seasonal capacity.

The model includes a minimum mission-coverage constraint and a maximum of 48 facings. It also applies a conservative transfer rate based on the lower bound of the test estimate. Store clusters receive different planograms because demand patterns differ.

Outcome

After 12 weeks, the pilot shows a 1.8% increase in category contribution, a 6% reduction in SKU count, and improved availability on core products. Revenue is approximately unchanged. The result is considered successful because the objective was profitable, reliable choice rather than revenue growth alone. The pilot is extended, with quarterly monitoring for customer complaints, competitor switching, and changes in segment demand.

Learning points

1. Low sales do not automatically mean low strategic value.
2. Delisting analysis must estimate where demand goes.
3. Space reallocation can be more valuable than listing reduction alone.
4. Store clusters are essential when preferences differ.
5. A controlled pilot is preferable to an irreversible national change.

10. Technology, AI, and governance

Assortment systems increasingly combine point-of-sale data, inventory records, electronic shelf labels, planogram software, and machine-learning forecasts. AI can identify substitution patterns, forecast demand, and recommend space allocations. Its value depends on data quality and interpretability.

Important governance practices include:

• maintain a documented data dictionary and model version;
• distinguish observed sales from estimated lost sales;
• record assumptions about transfer, elasticity, and incrementality;
• test recommendations across stores, customer segments, and time periods;
• monitor bias against small, specialist, local, or culturally important products;
• require human review for delisting decisions;
• protect customer-level data through aggregation, access controls, and lawful processing.

A black-box model that repeatedly removes products serving minority needs may improve a narrow productivity metric while creating discriminatory access outcomes. Retailers should include customer representation, accessibility, and regulatory obligations in the objective and constraints.

11. Limitations and best practices

Quantitative assortment optimization faces several limitations. Historical data reflect previous assortment decisions and may not reveal demand for products never listed. Prices and space are not randomly assigned. Promotions create confounding effects. Supplier-funded activity may distort margin comparisons. New products lack stable history. Online search and substitution behaviour may differ from physical-store behaviour.

Best practices are therefore to:

• use multiple productivity measures;
• correct, or at least flag, stock-out periods;
• separate base and promotional demand;
• estimate transfer through tests where possible;
• use conservative assumptions for irreversible decisions;
• review results by store cluster and customer mission;
• model integer and physical constraints;
• compare outcomes with a baseline and control group;
• pilot major changes before full rollout;
• monitor post-decision effects and revise the model.

The final decision should be documented as a transparent business case: what changed, why it changed, which assumptions were used, what risks remain, and how success will be measured.

Chapter Summary

Quantitative assortment optimization links customer needs with financial and operational performance. SKU productivity identifies how effectively products generate sales, margin, and space returns, while Pareto analysis reveals concentration without automatically justifying delisting. Demand transfer and substitutability are essential because removed demand may move to retained products, leave the retailer, or disappear. Incrementality distinguishes genuine category growth from cannibalization.

Space elasticity estimates the response of sales to allocated space, but it must be interpreted cautiously because space is often assigned to products that already sell well. Optimization models can represent listing, facing, budget, range, availability, and regulatory constraints. Scenario and sensitivity analysis expose the effect of uncertain assumptions. The most reliable practice combines quantitative modelling, controlled pilots, customer-mission analysis, governance, and managerial judgement.

Key Terms

SKU: A distinct product and inventory identity.

Assortment: The set of products offered within a category or store.

SKU productivity: Value or volume generated relative to the number of SKUs, space, inventory, or another resource.

Pareto analysis: Ranking items by contribution and examining cumulative concentration.

Long tail: The group of numerous low-volume products beyond the category’s core items.

Demand transfer: Movement of demand from a removed or changed SKU to another product.

Substitutability: The extent to which products can fulfil similar customer needs.

Incrementality: Sales or contribution that would not otherwise have occurred.

Cannibalization: Reduction in one product’s demand caused by another product in the same assortment.

Space elasticity: Percentage change in sales associated with a percentage change in allocated space.

Facing: The number of visible product units positioned across the front of a shelf.

Contribution margin: Sales less relevant variable costs.

GMROII: Gross margin generated relative to average inventory investment.

Rationalization: Deliberate reduction or redesign of an assortment to improve performance.

Mixed-integer optimization: An optimization model containing continuous and discrete decision variables.

Hard constraint: A condition that must be satisfied.

Soft constraint: A preference that may be relaxed at a measurable penalty.

Scenario analysis: Evaluation of decisions under alternative assumptions or conditions.

Sensitivity analysis: Testing how outputs change when an input parameter changes.

Mission coverage: Representation of the customer needs or shopping occasions served by an assortment.

Review Questions

Short Answer Questions

1. What is the difference between sales per SKU and sales per facing?
2. Why should Pareto analysis not be used as an automatic delisting rule?
3. Define demand transfer and explain why it matters in rationalization.
4. What does space elasticity measure?
5. Identify two hard constraints that may affect assortment optimization.

Long Answer Questions

1. Explain how a retailer should combine SKU productivity, margin, customer coverage, and operational complexity when evaluating an assortment.
2. Discuss the assumptions and limitations of using cross-price elasticity to estimate substitutability.
3. Describe a rigorous process for estimating the incrementality of a new SKU.
4. Explain how mixed-integer optimization can represent listing and facing decisions.
5. Evaluate the role of AI in assortment optimization, including data, governance, and ethical risks.

Analytical / Application Questions

1. A delisted SKU sells 10,000 units with contribution of 2 per unit. Seventy percent of demand transfers to a retained SKU contributing 2.40 per unit, while 20% is lost and 10% is uncertain. Ignoring other costs, calculate the contribution change from the observed transfer and lost demand, and state what additional evidence is required before delisting.
2. A category has 50 SKUs. The first 10 generate 65% of sales, but four of the remaining SKUs serve an important dietary mission. Design a rationalization approach that uses Pareto evidence without eliminating mission coverage.
3. A retailer can choose between a revenue-maximizing scenario and a slightly lower-revenue scenario with higher contribution, better availability, and greater customer coverage. Construct a decision framework and identify the sensitivity tests that should be performed before implementation.

Multiple Choice Questions with Answers

Question 1

Which measure directly evaluates sales relative to allocated shelf space?

A. Sales per facing

B. Units per supplier

C. Category penetration

D. Stock cover

Correct Answer: A. Sales per facing

Question 2

In Pareto analysis, cumulative contribution is calculated after SKUs are first ranked by which principle?

A. Alphabetical order

B. Descending contribution to the selected measure

C. Ascending price only

D. Random store order

Correct Answer: B. Descending contribution to the selected measure

Question 3

If 80% of a delisted SKU’s demand transfers to retained products, what does the remaining 20% represent most directly?

A. Guaranteed incremental demand

B. Demand that must be allocated to loss, postponement, or estimation error

C. Supplier margin

D. Shelf elasticity

Correct Answer: B. Demand that must be allocated to loss, postponement, or estimation error

Question 4

A positive cross-price elasticity commonly suggests that two products are:

A. Complements

B. Unrelated in all circumstances

C. Substitutes

D. Regulated goods

Correct Answer: C. Substitutes

Question 5

Space elasticity is defined as:

A. Change in margin divided by inventory

B. Percentage change in sales divided by percentage change in space

C. Space divided by SKU count

D. Units divided by price

Correct Answer: B. Percentage change in sales divided by percentage change in space

Question 6

If a retailer has a maximum of 100 listings, which constraint is appropriate when x_i equals 1 for a listed SKU?

A. ∑x_i ≥ 100

B. ∑x_i = 0

C. ∑x_i ≤ 100

D. ∑x_i × 100 = 1

Correct Answer: C. ∑x_i ≤ 100

Question 7

Why can a low-selling SKU still be retained in an optimized assortment?

A. It always has the highest margin

B. It may fulfil an important customer mission or provide strategic value

C. Low sales always indicate high incrementality

D. It requires no inventory

Correct Answer: B. It may fulfil an important customer mission or provide strategic value

Question 8

A controlled delisting test is primarily useful for estimating:

A. Product colour

B. Demand transfer and lost demand

C. Supplier ownership

D. Shelf dimensions only

Correct Answer: B. Demand transfer and lost demand

Question 9

If sales increase by 10% and space increases by 50%, the local space elasticity is:

A. 0.2

B. 2.0

C. 5.0

D. 60

Correct Answer: A. 0.2

Question 10

Which issue is a major governance risk in AI-supported assortment decisions?

A. Using documented assumptions

B. Testing recommendations across store clusters

C. Systematically disadvantaging specialist or minority-need products

D. Monitoring post-decision outcomes

Correct Answer: C. Systematically disadvantaging specialist or minority-need products