Prescriptive Analytics and Decision Optimization

Introduction

Descriptive analytics asks what has happened, diagnostic analytics asks why it happened, and predictive analytics asks what is likely to happen. Prescriptive analytics addresses the managerial question that follows: what should the organization do? It combines forecasts, business rules, optimization models, simulation, and risk analysis to recommend actions that best satisfy organizational objectives while respecting constraints.

Prescriptive analytics does not eliminate managerial judgment. Rather, it makes the assumptions, trade-offs, and consequences of decisions more explicit. A useful recommendation must be mathematically feasible, economically meaningful, operationally implementable, ethically acceptable, and robust to uncertainty.

Examples include deciding which products to manufacture, how much inventory to hold, which customers to target, how to assign employees to shifts, how to route deliveries, and whether to invest in capacity. These decisions generally involve scarce resources, competing objectives, and uncertain outcomes.

Learning Objectives

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

1. Distinguish predictive analytics from prescriptive analytics.
2. Formulate objectives, decision variables, constraints, and parameters.
3. Use decision trees to evaluate sequential decisions under uncertainty.
4. Formulate and interpret linear optimization models.
5. Explain shadow prices, slack, sensitivity analysis, and scenario analysis.
6. Apply simulation and risk analysis when uncertainty is important.
7. Evaluate optimization recommendations in ethical, governance, and implementation contexts.

From Prediction to Prescription

A prescriptive model can be represented conceptually as:

data → forecasts → decision model → feasible actions → evaluation → implementation and monitoring

A forecast is not itself a decision. For example, a demand forecast of 10,000 units does not determine the appropriate production quantity. Production may be limited by labor, materials, storage, cash, contractual obligations, or environmental targets. Prescriptive analytics connects the forecast to an action by representing those conditions explicitly.

A decision problem normally contains the following elements:

The feasible set is the collection of all decisions satisfying the constraints. The optimal solution is the feasible decision producing the best objective value according to the model. An optimal model solution is not necessarily the best real-world decision if the model omits important constraints or uses unreliable data.

Decision Trees for Sequential Decisions

Decision trees are useful when decisions occur in stages and uncertain events affect later choices. A square node represents a decision controlled by the organization; a circular node represents an uncertain event; branches represent alternatives or outcomes; and terminal nodes contain payoffs or costs.

The standard evaluation method is rollback analysis. Beginning at the end of the tree, calculate expected values at chance nodes and select the best alternative at decision nodes.

For uncertain outcomes with probabilities p₁, p₂, …, pₙ and payoffs V₁, V₂, …, Vₙ, the expected monetary value is:

Here pᵢ is the probability of outcome i, Vᵢ is its monetary payoff, and the probabilities must satisfy 0 ≤ pᵢ ≤ 1 and ∑pᵢ = 1. EMV is appropriate when probabilities are credible and the decision maker is willing to evaluate alternatives by their probability-weighted average payoff. It does not describe the dispersion or downside risk of outcomes.

Worked Example: Capacity Investment

A company may invest in a small or large facility. Demand may be high with probability 0.6 or low with probability 0.4. The payoffs are:

For the small facility:

For the large facility:

The alternatives have the same EMV. A risk-neutral decision maker is indifferent on expected value, but a risk-averse manager may prefer the small facility because its worst outcome is positive. The example illustrates why expected value should be supplemented by scenario and risk analysis.

If a decision is made after receiving perfect information about demand, the expected value with perfect information is:

EVPI is an upper bound on the value of any information system. In practice, information is rarely perfect and may have acquisition, delay, privacy, or implementation costs.

Linear Optimization

Linear programming (LP) is a central prescriptive method for allocating scarce resources. It applies when the objective and constraints can be represented as linear relationships.

A general maximization model is:

The variable xⱼ is decision variable j; cⱼ is its objective coefficient; aᵢⱼ is the amount of resource i used by one unit of activity j; bᵢ is the available amount of resource i; and Z is the objective value. The model assumes proportionality, additivity, divisibility where appropriate, and known coefficients. If decisions must be whole numbers, integer or mixed-integer programming is required.

Worked Example: Product Mix

A manufacturer produces products A and B. Each unit produces the following contribution:

Let x₁ be units of A and x₂ units of B. The model is:

Because product A uses more labor per unit and both products use one machine hour, labor is relatively valuable. At x₁ = 40 and x₂ = 40, machine capacity is fully used and labor use is 120 hours, which is infeasible. At x₁ = 20 and x₂ = 60, labor use is 100 hours and machine use is 80 hours, so the plan is feasible. Its contribution is:

To identify the optimum, examine the intersection of the binding resource constraints:

Subtracting the second equation from the first gives x₁ = 20, and therefore x₂ = 60. The solution satisfies both demand limits and is optimal for this model. It uses all 100 labor hours and all 80 machine hours.

The graphical method works conveniently for two variables, but professional applications commonly contain hundreds or millions of variables and constraints. Spreadsheet solvers, Python optimization libraries, R packages, and commercial solvers use algorithms such as the simplex method, interior-point methods, and branch-and-bound for integer models.

Model Formulation Discipline

A reliable formulation follows a clear sequence:

1. State the managerial question.
2. Define decision variables in operational units.
3. Establish a consistent time horizon.
4. Specify the objective and whether it is maximized or minimized.
5. Translate physical, financial, legal, and policy requirements into constraints.
6. Check units and signs.
7. Test the model with simple cases.
8. Validate the solution with domain experts.

Common errors include maximizing revenue when contribution margin is required, double-counting resources, omitting non-negativity, confusing demand with sales, and using monthly capacity with annual demand.

Sensitivity Analysis

An optimization solution is conditional on its inputs. Sensitivity analysis examines how the solution changes when coefficients or resource limits change.

Shadow Prices

The shadow price of a resource is the marginal improvement in the objective associated with one additional unit of that resource, within a valid range and holding other assumptions constant. In the product-mix example, a shadow price for labor might indicate the maximum additional contribution obtainable from one more labor hour. It is not necessarily the market wage and should not be interpreted outside the allowable range.

For a constraint with right-hand side bᵢ, the shadow price is conceptually:

where ∆Z is the change in optimal objective value and ∆bᵢ is the change in resource availability. This marginal interpretation applies only while the current optimal basis remains valid.

A constraint with unused capacity has positive slack:

A binding constraint has zero slack. In a maximization problem, a nonbinding resource commonly has a zero shadow price, although special cases such as degeneracy require careful interpretation.

Reduced Costs and Ranges

A reduced cost indicates how much an objective coefficient for a currently unused variable would need to improve before that variable becomes attractive in the current model. Solver reports may also provide allowable increases and decreases for objective coefficients and right-hand sides. These ranges are local sensitivity results, not guarantees under simultaneous large changes.

Sensitivity analysis supports managerial questions such as:

• Would overtime be worthwhile?
• What is the value of expanding warehouse capacity?
• How much can a supplier’s price increase before the sourcing plan changes?
• Which forecast assumptions deserve the most attention?

Scenario Analysis and Robust Decisions

Scenario analysis evaluates a small number of coherent alternative futures, such as low, baseline, and high demand. A scenario should change related inputs consistently. For example, a recession scenario may reduce demand, increase credit costs, and change product mix rather than changing only one number.

A decision can be optimal in one scenario and poor in another. Managers may therefore compare:

• objective value in each scenario;
• feasibility under each scenario;
• downside outcomes;
• operational reversibility;
• required contingency actions.

A robust decision may not maximize the objective in the baseline scenario, but may perform acceptably across a wide range of conditions. Robustness is particularly valuable when forecasts are uncertain and decisions are difficult to reverse.

Simulation and Risk Analysis

Optimization often treats uncertain inputs as fixed estimates. Simulation instead represents uncertainty explicitly and repeatedly evaluates the system. A Monte Carlo simulation samples random values from specified probability distributions, calculates outcomes, and summarizes the resulting distribution.

A basic simulation process is:

1. Identify uncertain inputs.
2. Specify probability distributions using historical data or expert judgment.
3. Generate random draws.
4. Calculate the outcome for each trial.
5. Repeat many times.
6. Examine means, percentiles, probabilities, and extreme outcomes.

If Y₁, …, Yₙ are simulated outcomes, the estimated mean is:

The empirical probability of exceeding a threshold T is:

These are estimates whose precision increases with the number of trials, but simulation cannot correct a poorly specified distribution or an incorrect model structure.

Important risk measures include variance, standard deviation, percentiles, and downside probability. For a sample of n outcomes, sample standard deviation is:

The units of s are the same as the outcome Y. A 5th percentile profit, for example, describes a low-outcome boundary: approximately 5% of simulated outcomes are below it. It is not the same as the probability of loss.

Simulation and optimization can be combined. A candidate plan may be generated by an optimization model and then stress-tested through simulation. Alternatively, simulation optimization can search for decisions that perform well under uncertain conditions, although computational and modeling complexity increases.

A Detailed Illustrative Case: FreshBite Distribution

Background

FreshBite, a fictional regional food distributor, supplies restaurants with fresh meal kits. It must decide weekly quantities of three kits: vegetarian, standard, and premium. Each kit uses packing labor and refrigerated storage. Contribution margins and resource requirements are estimated from recent operations.

Weekly availability is 1,000 labor hours and 5,000 storage units. Demand limits are 3,000 vegetarian, 2,500 standard, and 1,500 premium kits.

Problem Formulation

Let v, s, and p denote weekly quantities of vegetarian, standard, and premium kits. FreshBite seeks to maximize contribution:

The model assumes contribution margins are constant, all demand limits are known, kits are divisible for planning purposes, and no minimum service commitments or spoilage effects have been omitted.

Analysis and Solution

A solver would produce an optimal plan and report resource usage, slack, and sensitivity information. The analyst should not simply forward the solver output. The plan must be checked against delivery schedules, minimum customer commitments, labor shift patterns, and food-safety capacity. If the model recommends fractional kits, integer restrictions or practical rounding rules are required.

Suppose the solution recommends increasing premium production after a demand forecast revision. Management should ask whether the premium forecast is reliable, whether premium kits have higher spoilage, and whether the plan creates excessive concentration in one product. A scenario analysis might compare baseline demand with a 20% premium-demand reduction and a 15% labor shortage.

Risk and Governance

FreshBite should simulate demand and spoilage, then report expected contribution, the probability of stockout, the probability of waste exceeding a threshold, and the 10th percentile contribution. Forecast distributions should be documented, not selected merely to make the recommendation attractive.

The final decision may include a contingent policy: produce a baseline quantity, reserve labor for late-week replenishment, and trigger additional production only when confirmed orders exceed a threshold. This policy may be less mathematically elegant than a single fixed plan but more resilient operationally.

Outcome and Learning Points

The case demonstrates that optimization is a structured decision aid rather than an automatic answer. Good practice requires:

• a clearly defined objective;
• measurable and complete constraints;
• validation of data and assumptions;
• sensitivity and scenario analysis;
• communication of uncertainty;
• monitoring after implementation.

Multiobjective Decisions and Trade-offs

Business decisions frequently involve several objectives: profit, service quality, employee fairness, emissions, resilience, and customer retention. These objectives may conflict. A weighted objective can represent preferences:

Here fᵢ(x) is objective i, wᵢ is its weight, and weights should reflect normalized or economically comparable measures. Arbitrary weights can conceal value judgments. An alternative is to treat important objectives as constraints, such as requiring a minimum service level or maximum emissions. Analysts should present trade-offs rather than implying that a single score is objectively correct.

Ethical, Legal, and Governance Considerations

Prescriptive recommendations can affect employees, customers, suppliers, and communities. Key governance practices include:

• Transparency: document data sources, assumptions, objective functions, and constraints.
• Accountability: assign responsibility for approving and reviewing decisions.
• Fairness: test whether allocations or scores disadvantage protected groups.
• Privacy: use only necessary personal data and protect sensitive information.
• Auditability: retain model versions, input data, solver settings, and decision logs.
• Human oversight: permit review and override when circumstances are unusual.
• Security: control access to models and prevent manipulation of inputs.

An apparently efficient model may reproduce historical discrimination, exploit workers, or shift environmental costs to vulnerable communities. Compliance with privacy, employment, consumer-protection, financial, and sector-specific regulation must be considered before deployment. Explainability is especially important when recommendations affect eligibility, pricing, credit, employment, or access to essential services.

Technology and Contemporary Practice

Modern analytics platforms connect data warehouses, forecasting systems, optimization solvers, dashboards, and workflow tools. Cloud computing makes large simulations and scenario analysis more accessible. Machine learning can improve demand forecasts, but a more accurate forecast does not guarantee a better decision if the optimization model is incomplete.

Generative AI can assist with model documentation, code drafts, and scenario descriptions, but it should not be trusted to invent constraints, validate arithmetic, or make unaudited high-impact decisions. Analysts must independently verify generated formulas, data transformations, and solver results. Model risk management increasingly emphasizes version control, reproducibility, monitoring for drift, and documented human approval.

Challenges and Limitations

Prescriptive analytics faces several recurring limitations:

1. Data quality: inaccurate costs, demand, or capacity produce unreliable recommendations.
2. Model misspecification: omitted constraints can make an apparently optimal plan infeasible.
3. Parameter uncertainty: estimated coefficients may be more uncertain than the solver output suggests.
4. Structural change: historical relationships may fail during disruption or market change.
5. Overoptimization: maximizing a narrow metric can damage long-term performance.
6. Computational complexity: integer, nonlinear, and large-scale models may require approximations.
7. Implementation resistance: users may reject opaque recommendations or lack authority to act.
8. False precision: a result such as 37.42% can appear more certain than the evidence supports.

Best practice is to combine mathematical rigor with domain expertise, scenario testing, clear communication, and post-implementation learning.

Chapter Summary

Prescriptive analytics transforms information into recommended action by combining objectives, decision variables, constraints, forecasts, and uncertainty analysis. Decision trees evaluate sequential choices using rollback and expected monetary value, while recognizing that risk preferences matter. Linear programming allocates scarce resources under linear assumptions; its solutions can be examined through slack, shadow prices, reduced costs, and sensitivity ranges. Scenario analysis tests coherent alternative futures, and simulation estimates distributions of outcomes when demand, costs, or other inputs are uncertain. Multiobjective models make trade-offs explicit, but weights and priorities must be governed transparently. Effective practice also requires data validation, ethical review, human oversight, implementation planning, and continuous monitoring.

Key Terms

Prescriptive analytics: Analytics that recommends actions to achieve specified objectives under constraints.

Decision variable: A controllable quantity or choice represented in a model.

Parameter: An input value treated as known or estimated in a model.

Objective function: A mathematical expression to maximize or minimize.

Constraint: A restriction defining which decisions are feasible.

Feasible solution: A decision satisfying all model constraints.

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

Decision tree: A branching representation of sequential decisions and uncertain events.

Expected monetary value: Probability-weighted average payoff across possible outcomes.

EVPI: Expected value of perfect information.

Linear programming: Optimization with linear objective and constraint relationships.

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

Slack: Unused amount of a constrained resource.

Reduced cost: Objective-coefficient improvement required for an inactive variable to become attractive, under local conditions.

Sensitivity analysis: Study of how model results respond to input changes.

Scenario analysis: Evaluation of decisions under coherent alternative futures.

Monte Carlo simulation: Repeated random sampling used to estimate an outcome distribution.

Robust decision: A decision that performs acceptably across varied conditions.

Model governance: Processes ensuring that models are documented, controlled, reviewed, and responsibly used.

Review Questions

Short Answer Questions

1. How does prescriptive analytics differ from predictive analytics?
2. What are the four essential components of a linear optimization model?
3. Why might a manager reject the alternative with the highest expected monetary value?
4. What does a shadow price represent?
5. Why is simulation useful when model inputs are uncertain?

Long Answer Questions

1. Explain the process of formulating a business problem as a linear programming model. Include decision variables, objective function, constraints, and assumptions.
2. Describe how a decision tree is constructed and evaluated using rollback analysis. Discuss the limitations of expected monetary value.
3. Compare sensitivity analysis, scenario analysis, and Monte Carlo simulation.
4. Explain how multiobjective optimization can represent trade-offs among profit, service, fairness, and sustainability.
5. Discuss the ethical and governance controls required before deploying a prescriptive analytics system.

Analytical / Application Questions

1. A retailer has limited labor and shelf space and must choose quantities of three products. Formulate an LP model, identify likely binding constraints, and explain what information a shadow-price report would provide.
2. A firm is choosing between two investments with different high- and low-demand payoffs. Construct a decision tree, calculate EMV for each alternative, and recommend additional risk measures.
3. Design a simulation study for a delivery company facing uncertain demand, travel time, and fuel cost. Specify inputs, outputs, performance measures, and a policy for acting on the results.

Multiple Choice Questions with Answers

Question 1

Which statement best describes prescriptive analytics?

A. It reports historical sales totals without interpretation.

B. It recommends actions subject to objectives and constraints.

C. It records raw transactions in a database.

D. It forecasts a variable without considering decisions.

Correct Answer: B. It recommends actions subject to objectives and constraints.

Question 2

At a chance node in a decision tree, the usual rollback operation is to:

A. Select the branch with the highest payoff regardless of probability.

B. Average all payoffs using equal weights.

C. Calculate the probability-weighted expected value.

D. Remove the branch with the lowest probability.

Correct Answer: C. Calculate the probability-weighted expected value.

Question 3

In a linear optimization model, xⱼ most commonly represents:

A. A decision variable.

B. A probability distribution.

C. A solver error.

D. A historical observation.

Correct Answer: A. A decision variable.

Question 4

A resource constraint has 20 units of unused capacity. Its slack is:

A. −20 units.

B. 0 units.

C. 10 units.

D. 20 units.

Correct Answer: D. 20 units.

Question 5

If an alternative has payoffs of $100,000 with probability 0.7 and $20,000 with probability 0.3, its EMV is:

A. $64,000.

B. $76,000.

C. $84,000.

D. $120,000.

Correct Answer: C. $84,000.

Question 6

A shadow price is best interpreted as:

A. The total accounting cost of all resources.

B. The marginal objective improvement from relaxing a constraint within a valid range.

C. The probability that a constraint will bind.

D. The selling price of the final product.

Correct Answer: B. The marginal objective improvement from relaxing a constraint within a valid range.

Question 7

What is the principal purpose of scenario analysis?

A. To test decisions under coherent alternative futures.

B. To guarantee that the baseline forecast is correct.

C. To replace all historical data with random values.

D. To eliminate the need for constraints.

Correct Answer: A. To test decisions under coherent alternative futures.

Question 8

Monte Carlo simulation primarily produces:

A. A single guaranteed outcome.

B. A distribution of possible outcomes based on repeated sampling.

C. A legal contract for decision implementation.

D. A proof that the optimization model is correct.

Correct Answer: B. A distribution of possible outcomes based on repeated sampling.

Question 9

Which condition most strongly suggests using integer programming rather than ordinary linear programming?

A. The objective is to maximize profit.

B. Resource usage is measured in hours.

C. The decision must involve whole vehicles rather than fractional vehicles.

D. Demand is uncertain.

Correct Answer: C. The decision must involve whole vehicles rather than fractional vehicles.

Question 10

Why should optimization results be reviewed by domain experts?

A. Solvers cannot perform arithmetic.

B. Experts can identify omitted constraints and implementation problems.

C. Experts always produce a higher objective value.

D. Optimization models never use data.

Correct Answer: B. Experts can identify omitted constraints and implementation problems.