Prescriptive Analytics and Optimization

Introduction

Descriptive analytics explains what has happened, diagnostic analytics investigates why it happened, and predictive analytics estimates what is likely to happen. Prescriptive analytics addresses the next managerial question: What should be done? It combines data, business rules, mathematical models, forecasts, and alternative scenarios to recommend actions.

Prescriptive analytics is particularly valuable when an organisation must allocate scarce resources. Examples include deciding which products to manufacture, how many employees to schedule, which deliveries to route, how much inventory to hold, or which marketing channels to fund. In each case, decision-makers must balance an objective—such as profit, service quality, or cost—with limitations such as budget, capacity, labour availability, or regulatory requirements.

The purpose of optimization is not to replace managerial judgement. Rather, it provides a transparent and disciplined way to compare feasible alternatives. A model can identify a mathematically attractive solution, while managers assess issues that are difficult to quantify, such as employee morale, brand reputation, resilience, and ethical impact.

Learning Objectives

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

1. Explain the components of a prescriptive decision model.
2. Distinguish decision variables, parameters, objective functions, and constraints.
3. Formulate and solve a basic linear programming problem.
4. Interpret feasible solutions, binding constraints, and slack.
5. Use scenario analysis and sensitivity analysis to assess uncertainty.
6. Explain how simulation supports decisions when uncertainty is substantial.
7. Evaluate the managerial, ethical, and practical limitations of optimization models.

1. The Structure of a Decision Model

A decision model translates a business problem into a formal representation. A useful model normally contains five elements:

A model begins with a decision statement. For example: “How many units of each product should the factory produce next week to maximize contribution, given available labour and machine capacity?” This statement identifies the decision, the objective, and relevant restrictions.

Decision variables

Decision variables represent controllable choices. If x₁ is the number of units of product A and x₂ is the number of units of product B, then the model’s solution determines the values of x₁ and x₂. Variables may be continuous, integer, or binary. A continuous variable can take values such as 12.5; an integer variable must be a whole number; a binary variable takes only 0 or 1 and often represents a yes/no decision.

Parameters and assumptions

Parameters are inputs treated as fixed during analysis. They may come from accounting records, operational databases, forecasts, engineering estimates, or expert judgement. Because parameters are often uncertain, analysts should record their source, date, units, and confidence level.

Every model includes assumptions. For example, a linear programming model may assume that unit profit is constant, resources are divisible, and product requirements do not change with production volume. Assumptions make a model manageable, but they also determine when its results are credible.

Objective functions

An objective function expresses the goal mathematically. A general maximization objective can be written as:

Here, Z is the total objective value, x_j is the level of decision variable j, c_j is the contribution of one unit of variable j, and n is the number of decision variables. The expression is applicable when total contribution is additive and proportional to activity levels. For cost minimization, the objective is instead Minimize Z = ∑(j=1 to n) c_jx_j, where c_j represents unit cost.

Unit consistency is essential. If c_j is measured in dollars per unit and x_j in units, then each product c_jx_j is measured in dollars.

Constraints

Constraints define the feasible region—the set of decisions that satisfy all stated conditions. A general resource constraint is:

In this expression, a_ij is the amount of resource i used by one unit of activity j, b_i is the available amount of resource i, and m is the number of constraints. The inequality applies when usage cannot exceed availability. Other relationships may use ≥, for example a minimum staffing requirement, or =, for example a balance requirement.

Non-negativity constraints are common:

They prevent the model from recommending negative production, staffing, or shipments. Integer restrictions are added when fractional decisions are not meaningful:

2. Linear Programming

Linear programming (LP) is an optimization method in which the objective and constraints are linear functions of the decision variables. LP is appropriate when proportionality, additivity, divisibility, and known coefficients are reasonable assumptions.

A standard LP model is:

The feasible region is formed by the intersection of all constraints. For a two-variable problem, it can be plotted on a graph. A fundamental property of LP is that, when an optimum exists, at least one optimal solution occurs at an extreme point (corner point) of the feasible region. This supports the graphical method for small problems, although software is normally used for larger models.

Worked example: product mix

A small manufacturer produces Products A and B. The contribution margin is $40 per unit of A and $30 per unit of B. Each unit requires the following resources:

Let x be units of A and y be units of B. The model is:

The intercepts and intersection points provide candidate extreme points. The resource constraints intersect when:

From the first equation, y = 100 - 2x. Substitution into the second gives x + 2(100 - 2x) = 80, so -3x = -120 and x = 40. Therefore y = 20.

The candidate points and objective values are:

The optimal solution is to produce 40 units of A and 20 units of B, generating a contribution of $2,200. Both resource constraints bind because all 100 labour hours and all 80 machine hours are used.

This result is valid only under the model assumptions. If demand for A is limited to 25 units, a further constraint x ≤ 25 must be added. If products must be produced in whole units, an integer programming model may be more appropriate.

Feasibility, optimality, and unboundedness

A model is feasible if at least one decision satisfies every constraint. It is infeasible if constraints conflict—for example, a requirement to use at least 100 hours when only 80 hours are available. A model is unbounded when the objective can improve indefinitely, usually because a necessary limiting constraint has been omitted. These outcomes are modelling warnings, not useful business recommendations.

3. Interpreting an Optimization Solution

An optimization report should be read as more than a list of variable values.

Binding constraints are satisfied as equalities at the solution. They identify resources that are fully used or requirements that are exactly met. Slack is unused capacity. For a constraint of the form resource used ≤ resource available, slack is:

Slack has the same units as the relevant resource. In the product-mix example, the optimal solution has zero slack for both labour and machine hours.

A shadow price (or dual value) estimates the change in the optimal objective value resulting from a one-unit increase in the right-hand side of a constraint, within an allowable range and assuming other parameters remain unchanged. Shadow prices support decisions such as whether to purchase additional capacity. They should not be interpreted as universally valid prices outside the sensitivity range.

Optimization can also involve multiple optimal solutions. This occurs when the objective function is parallel to a binding boundary. In such cases, managers may choose among equally profitable alternatives using secondary criteria such as risk, flexibility, or customer priority.

4. Scenario Analysis

Scenario analysis evaluates the model under alternative assumptions. It is especially useful when the future is uncertain but managers can describe plausible states, such as low, expected, and high demand.

A scenario table might include:

Scenario analysis changes one or more inputs and resolves the model. It differs from sensitivity analysis, which usually studies the effect of changing one parameter over a range while holding other inputs constant. Scenario analysis is often more realistic when several variables change together.

A good scenario analysis should identify:

• the assumptions changed;
• whether the model remains feasible;
• the recommended decision;
• the objective value;
• operational risks and implementation requirements.

Scenario analysis should not be confused with probability forecasting. Scenarios may be plausible narratives without assigned probabilities. If probabilities are available, expected value can be calculated, but the analyst should also examine downside outcomes.

5. Sensitivity Analysis

Sensitivity analysis asks how strongly the recommendation depends on uncertain inputs. Important inputs include prices, costs, demand, capacity, processing times, and forecast errors.

In an LP model, the allowable range for an objective coefficient indicates how far that coefficient may change before the current optimal basis changes. Similarly, the allowable range for a constraint’s right-hand side indicates how far resource availability may change before the associated shadow price no longer applies. These ranges are model-specific and should be obtained from the solver output.

For example, if a product’s contribution rises slightly but the optimal production mix remains unchanged, the decision is relatively stable with respect to that contribution. If a small change causes a different product mix, the decision is sensitive and deserves managerial attention.

Sensitivity analysis can be conducted through:

1. One-way analysis: change one input at a time.
2. Two-way analysis: vary two inputs and display outcomes in a table or chart.
3. Break-even analysis: identify the input value at which two alternatives become equally attractive.
4. Stress testing: examine deliberately adverse conditions.

Sensitivity analysis does not prove that a forecast is correct. It reveals the consequences of being wrong and helps prioritise data-quality improvements.

6. Simulation for Uncertain Operations

Optimization often assumes known coefficients, whereas simulation represents uncertainty explicitly. Simulation repeatedly generates possible outcomes using probability distributions and records performance measures. It is useful when systems include queues, random arrivals, variable service times, stockouts, failures, or complex interactions that are difficult to express in a simple LP model.

A basic Monte Carlo procedure is:

1. Define uncertain inputs and their probability distributions.
2. Generate random values for one trial.
3. Apply the operational rules or decision policy.
4. Record outcomes such as cost, profit, waiting time, or service level.
5. Repeat many trials.
6. Summarize the distribution of outcomes.

If Y_r is the outcome of trial r and there are N trials, the estimated mean outcome is:

The result is an estimate, not an exact population value. More trials generally reduce random simulation error, but they do not correct a badly specified model.

Simulation and optimization answer different questions. Simulation evaluates a proposed policy under uncertainty; optimization searches for a policy that performs best according to an objective. They can be combined: an optimization model can generate candidate decisions, while simulation tests their robustness under uncertain demand and operating conditions.

For example, an inventory manager may optimize the order quantity using expected demand, then simulate thousands of demand outcomes to estimate stockout probability, holding cost, and service level. A slightly less profitable policy may be preferred if it substantially reduces the risk of stockouts.

7. Applied Case: Regional Meal-Kit Allocation

This is an illustrative case. It demonstrates how prescriptive analytics can support a practical resource-allocation decision.

Background

FreshBox prepares vegetarian and family meal kits for regional delivery. Each week, management decides how many kits to prepare. The vegetarian kit contributes $18, while the family kit contributes $24. The preparation team has 600 labour hours. Each vegetarian kit requires 0.5 labour hours and each family kit requires 0.75 labour hours. Refrigerated storage provides 500 kit-hours, with each vegetarian kit using 0.4 kit-hours and each family kit using 0.8 kit-hours. Weekly demand is at most 800 vegetarian kits and 500 family kits.

Problem formulation

Let v be vegetarian kits and f be family kits. The model is:

The coefficients have consistent units: labour coefficients are hours per kit, storage coefficients are kit-hours per kit, and the objective coefficients are dollars per kit.

Analysis and approach

A solver would identify the best feasible combination while respecting demand limits. Management should then examine whether the solution is operationally credible. For example, a recommendation near the full demand limit may require reliable ingredient supply and accurate demand forecasts. If the variables must be whole kits, integer restrictions should be added, although the difference may be negligible at this scale.

The team should also test scenarios: a 20% fall in family-kit demand, a 10% reduction in labour availability, and an increase in family-kit contribution due to a price change. Simulation can estimate the probability that actual demand exceeds production and the resulting lost-margin cost.

Outcome and decision use

The final decision should not be based only on the highest expected contribution. Management might choose a solution with slightly lower expected profit if it uses less storage, protects service levels, or remains effective across several demand scenarios. The model therefore acts as a decision-support instrument rather than an automatic order.

Learning points

1. Business language must be converted into explicit variables, units, and constraints.
2. Demand limits are as important as production capacity.
3. Scenario and sensitivity analysis reveal whether a solution is robust.
4. Operational and ethical considerations may justify rejecting the purely mathematical optimum.

8. Implementation, Governance, and Ethics

A prescriptive model is only as reliable as its data, assumptions, and implementation process. Best practice includes:

• documenting every parameter and its source;
• validating units and data ranges;
• checking whether historical data reflect current conditions;
• testing feasibility before interpreting optimality;
• separating training, testing, and live decision environments where predictive inputs are used;
• retaining model versions and decision logs;
• assigning responsibility for approving model changes;
• monitoring actual outcomes against recommended outcomes.

Optimization can reproduce or intensify bias. A staff-scheduling model that minimizes cost may systematically assign undesirable shifts to a particular group if historical availability data reflect unequal treatment. A credit or pricing model may disadvantage protected groups through apparently neutral variables. Analysts should examine fairness, explainability, accessibility, and the consequences of errors.

Privacy and security also matter. Models may use employee, customer, location, or transaction data. Data collection should have a legitimate purpose, appropriate access controls, retention limits, and compliance with applicable privacy law. A recommendation should be explainable enough for affected decision-makers to understand its principal drivers.

Automation requires human oversight. Important decisions should have escalation rules, exception handling, and a documented process for challenging a recommendation. Artificial intelligence may assist with forecasting, feature generation, or large-scale search, but it does not remove the need to define objectives, constraints, and governance standards. An opaque AI recommendation can be less accountable than a transparent mathematical model.

9. Limitations and Contemporary Practice

Optimization models can fail for several reasons. A model may omit an important constraint, use unreliable forecasts, assume constant unit costs when economies of scale exist, or optimise a measurable target while neglecting quality. A mathematically optimal solution may also be infeasible in practice because of changeover times, supplier reliability, employee acceptance, or implementation delays.

Contemporary analytics teams increasingly combine optimization with cloud data platforms, real-time operational data, digital twins, and machine-learning forecasts. These technologies can improve responsiveness, but they also increase model complexity and governance requirements. Re-optimization at high frequency may create unstable plans unless changes are subject to minimum run lengths, review thresholds, or service commitments.

A disciplined workflow is therefore:

1. Define the decision and the accountable owner.
2. Identify objectives, constraints, and non-negotiable rules.
3. Collect and validate data.
4. Formulate the simplest adequate model.
5. Verify logic using extreme and known cases.
6. Solve and interpret the result.
7. Test scenarios, sensitivity, and uncertainty.
8. Review ethical and operational implications.
9. Implement with monitoring and human oversight.

Chapter Summary

Prescriptive analytics recommends actions by combining data, objectives, constraints, and decision rules. A decision model contains decision variables, parameters, an objective function, constraints, and a solution. Linear programming is appropriate when relationships are linear and coefficients are sufficiently known. Its results must be interpreted through feasibility, binding constraints, slack, and sensitivity information—not merely through the recommended variable values.

Scenario analysis evaluates coherent alternative futures, while sensitivity analysis examines how results respond to changes in individual inputs or ranges. Simulation represents uncertainty by repeatedly generating possible outcomes and is particularly valuable for queues, inventory, demand, and other stochastic operations. Optimization and simulation can be combined to find decisions that are both effective and robust.

Professional use requires careful formulation, validation, documentation, ethical review, privacy protection, and human oversight. The best prescriptive models improve managerial judgement without pretending that every important consideration can be expressed as a number.

Key Terms

Decision variable: A controllable quantity whose value is selected by the model.

Parameter: An input treated as fixed or estimated for a particular analysis.

Objective function: A mathematical expression representing the goal to maximize or minimize.

Constraint: A condition that limits allowable decisions.

Feasible solution: A decision satisfying every model constraint.

Optimal solution: A feasible solution with the best objective value under the model.

Linear programming: Optimization with linear objective and constraint relationships.

Integer programming: Optimization requiring one or more decision variables to be whole numbers.

Binary variable: A variable restricted to 0 or 1.

Feasible region: The set of all feasible combinations of decision variables.

Binding constraint: A constraint satisfied exactly at the solution.

Slack: Unused amount of a resource or requirement allowance.

Shadow price: Marginal value of relaxing a constraint within a valid sensitivity range.

Scenario analysis: Evaluation of decisions under alternative sets of assumptions.

Sensitivity analysis: Study of how results change when model inputs vary.

Simulation: Repeated representation of uncertain system behaviour.

Monte Carlo simulation: Simulation using repeated random sampling from probability distributions.

Prescriptive analytics: Analytics that recommends actions or decisions.

Model governance: Processes for documenting, validating, monitoring, and controlling analytical models.

Robustness: The extent to which a recommendation remains satisfactory under changing conditions.

Review Questions

Short Answer Questions

1. What are the five principal components of a prescriptive decision model?
2. What is the difference between a binding constraint and slack?
3. State two assumptions commonly associated with linear programming.
4. How does scenario analysis differ from sensitivity analysis?
5. Why might simulation be useful for inventory decisions?

Long Answer Questions

1. Explain how a business problem can be translated into a linear programming model. Use decision variables, parameters, an objective function, and constraints in your explanation.
2. Discuss the meaning and managerial use of shadow prices, including why their interpretation is limited to an allowable range.
3. Compare optimization and simulation as decision-support methods. Explain how they can be used together.
4. Describe a systematic workflow for building, validating, implementing, and monitoring a prescriptive model.
5. Evaluate the ethical risks of using optimization for employee scheduling or customer allocation.

Analytical / Application Questions

1. A firm earns $50 contribution per unit of Product X and $35 per unit of Product Y. X requires 3 labour hours and 1 machine hour; Y requires 2 labour hours and 2 machine hours. Weekly availability is 120 labour hours and 80 machine hours. Formulate the LP model and identify the production plan by evaluating the feasible corner points.
2. An inventory model recommends a policy based on expected demand. Design a scenario and simulation analysis to assess stockout risk, holding cost, and service level when demand is uncertain.
3. An optimization model recommends assigning all available overtime to one department. Explain what data, sensitivity results, operational constraints, and ethical considerations should be reviewed before implementing the recommendation.

Multiple Choice Questions with Answers

Question 1

Which component of a prescriptive model represents the choices controlled by a decision-maker?

A. Decision variables

B. Parameters

C. Constraints

D. Historical records

Correct Answer: A. Decision variables

Question 2

In a maximization model, what does a binding resource constraint indicate?

A. The resource is unused

B. The resource is fully used at the solution

C. The model is infeasible

D. The objective is minimized

Correct Answer: B. The resource is fully used at the solution

Question 3

Which condition is most appropriate for linear programming?

A. All relationships are unknown

B. Objective and constraint relationships are reasonably linear

C. All decisions must be random

D. No parameters are available

Correct Answer: B. Objective and constraint relationships are reasonably linear

Question 4

In the product-mix example, what is the optimal production plan?

A. 0 units of A and 40 units of B

B. 40 units of A and 20 units of B

C. 50 units of A and 0 units of B

D. 20 units of A and 40 units of B

Correct Answer: B. 40 units of A and 20 units of B

Question 5

What does slack measure in a constraint of the form resource used ≤ resource available?

A. The objective value

B. The unused amount of the resource

C. The number of decision variables

D. The forecast error

Correct Answer: B. The unused amount of the resource

Question 6

What is the principal purpose of sensitivity analysis?

A. To remove all uncertainty from the data

B. To examine how results respond to changes in inputs

C. To replace constraints with objectives

D. To guarantee a feasible solution

Correct Answer: B. To examine how results respond to changes in inputs

Question 7

Which method repeatedly generates random outcomes to represent uncertainty?

A. Linear programming

B. Scenario naming

C. Monte Carlo simulation

D. Constraint screening

Correct Answer: C. Monte Carlo simulation

Question 8

A binary decision variable normally takes which values?

A. Any positive real number

B. Only 0 or 1

C. Only negative values

D. Any whole number above 10

Correct Answer: B. Only 0 or 1

Question 9

Why should a shadow price not be applied without qualification to a very large capacity change?

A. It is measured in the wrong units

B. It is valid only within a model-specific sensitivity range

C. It always represents a fixed market price

D. It applies only to binary variables

Correct Answer: B. It is valid only within a model-specific sensitivity range

Question 10

Which practice most improves responsible use of an automated optimization model?

A. Removing all human review

B. Ignoring data provenance

C. Documenting assumptions and monitoring outcomes

D. Using the most complex model available

Correct Answer: C. Documenting assumptions and monitoring outcomes