Prescriptive Analytics, Optimization and Simulation

Introduction

Descriptive analytics asks what happened; diagnostic analytics asks why; predictive analytics estimates what may happen; prescriptive analytics addresses what should be done. Prescriptive analytics combines data, models, organizational objectives, constraints, uncertainty, and managerial judgment to recommend actions.

In business analytics, prescriptive models support decisions such as product mix, staffing, transportation, inventory, project selection, pricing, and capacity planning. The central idea is to translate a decision problem into a formal model containing:

• Decision variables: controllable choices.
• Objective function: the quantity to maximize or minimize.
• Constraints: limits, requirements, or policy conditions.
• Parameters: known or estimated inputs.
• Scenarios: alternative assumptions about uncertain conditions.

A mathematically optimal solution is not automatically a good managerial decision. Its value depends on data quality, realistic assumptions, implementability, fairness, and alignment with organizational strategy.

Learning Objectives

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

1. Formulate decision problems using variables, objectives, constraints, and scenarios.
2. Solve and interpret basic linear optimization models.
3. Explain feasible regions, binding constraints, shadow prices, and sensitivity analysis.
4. Use simulation to represent uncertainty and estimate the distribution of outcomes.
5. Construct and evaluate decision trees using expected monetary value.
6. Compare optimization, simulation, and decision-tree approaches.
7. Evaluate ethical, governance, and implementation issues in prescriptive analytics.

1. From Prediction to Prescription

A prediction estimates an uncertain outcome. A prescription recommends an action, often by combining predictions with costs, benefits, and restrictions. For example, a demand forecast may estimate next month’s demand for three products. An optimization model can then determine how much to produce, subject to labor and material availability.

Prescriptive analytics generally follows this workflow:

1. Define the decision and decision maker.
2. Identify feasible actions and relevant outcomes.
3. Gather and validate data.
4. Formulate the objective and constraints.
5. Solve the model or simulate alternatives.
6. Test sensitivity to assumptions.
7. Communicate the recommendation and its risks.
8. Monitor actual results and update the model.

The workflow is iterative. A model that cannot be explained, audited, or implemented should be revised, even if its numerical solution is optimal.

2. Mathematical Formulation of Decision Models

2.1 Decision variables and parameters

A decision variable represents an action. If a manufacturer chooses production quantities, x₁ may represent units of product 1 and x₂ units of product 2. A parameter is an input that is treated as known within a model, such as profit per unit or available labor hours.

Variables may be continuous, integer, or binary. Continuous variables can take values such as 2.5. Integer variables must be whole numbers, while binary variables take values 0 or 1 and often represent yes/no decisions.

2.2 Objective functions

An objective function measures the value of a decision. A general single-objective model can be written as:

Here, Z is the objective value and x_j are decision variables. In a linear model:

where c_j is the contribution, cost, or benefit per unit of variable x_j. Units must be consistent: if c_j is dollars per unit and x_j is units, Z is dollars.

Common objectives include maximizing profit, contribution margin, service coverage, or customer value, and minimizing cost, travel time, emissions, or risk. Multiple objectives may be handled through priorities, constraints, or a weighted objective such as:

where w₁ and w₂ are nonnegative weights. Weighted objectives require careful normalization because profit, emissions, and service levels may have different units and scales.

2.3 Constraints

Constraints define the feasible set of actions. A resource constraint commonly has the form:

where a_ij is the amount of resource i used by one unit of activity j, and b_i is the available amount of resource i. The units on both sides must match.

Other forms include:

• ∑ a_ijx_j ≥ b_i for minimum requirements.
• ∑ a_ijx_j = b_i for exact balance conditions.
• x_j ≥ 0 for non-negativity.
• x_j ∈ Z for integer decisions.
• y ∈ {0,1} for binary decisions.

Constraints may represent physical capacity, budgets, legal requirements, contractual commitments, staffing rules, or service standards.

2.4 Feasibility and optimality

A solution is feasible if it satisfies every constraint. The feasible region is the set of all feasible solutions. An optimal solution is a feasible solution with the best objective value according to the stated objective.

A model can be infeasible when requirements conflict. It can be unbounded when the objective can improve indefinitely without violating constraints. Both conditions usually indicate a formulation or data problem rather than a useful business result.

3. Linear Programming

A linear programming (LP) model has a linear objective and linear constraints. Its standard structure is:

Here, x is the vector of decision variables, c contains objective coefficients, A contains constraint coefficients, and b contains resource limits. LP assumes proportionality, additivity, divisibility, and known coefficients. These assumptions should be examined rather than accepted automatically.

Worked Example: Product Mix

A small factory produces tables and chairs. Each table contributes $80 and uses 4 labor hours and 3 units of wood. Each chair contributes $30 and uses 1 labor hour and 1 unit of wood. The factory has 40 labor hours and 24 units of wood available. Let T be tables and C be chairs.

At the intersection of the two resource constraints:

Subtracting the second equation from the first gives T = 16. Substitution gives C = -24, which is infeasible. Thus, the constraint lines intersect outside the nonnegative region. Candidate corner points are therefore found from the axes and relevant intersections.

On the T-axis, wood permits T = 8, producing profit $640. On the C-axis, labor permits C = 24, producing profit $720. The labor and wood constraints are not simultaneously binding at a nonnegative intersection because tables consume relatively more labor than wood. The optimal feasible corner is consequently (T,C) = (0,24), with maximum contribution $720.

This result does not mean tables are undesirable in every circumstance. It means that, under the stated contributions and resource limits, replacing chairs with tables reduces contribution. A solver would confirm the corner-point analysis for this two-variable LP.

Graphical interpretation

For two variables, constraints can be plotted as lines and the feasible region identified. For an LP, at least one optimal solution occurs at an extreme point of the feasible region, provided an optimum exists. Commercial and open-source solvers extend this principle computationally using algorithms such as the simplex method and interior-point methods.

4. Integer and Mixed-Integer Optimization

LP may recommend fractional decisions that are impossible or inappropriate, such as 3.7 delivery vehicles or 0.4 retail stores. Integer programming requires selected variables to be integers. Mixed-integer linear programming (MILP) combines continuous and integer variables.

Binary variables are useful for fixed-charge decisions. If opening a facility costs F and y equals 1 when it is opened, the objective may include -Fy. A linking constraint can be written as:

where M is a valid upper bound on activity x. If y = 0, then x = 0; if y = 1, activity may occur up to M. The value of M should be as tight as practical, because an unnecessarily large value can make computation slower and weaken numerical reliability.

Integer models are often computationally harder than LP models. A rounded LP solution may violate constraints or be substantially worse than the true integer optimum; therefore, rounding should not be treated as a general solution method.

5. Sensitivity Analysis and Scenario Planning

A model’s recommendation depends on its inputs. Sensitivity analysis asks how the solution or objective changes when one or more parameters change.

5.1 Binding constraints and slack

The slack of a “less-than-or-equal-to” resource constraint is:

Slack has the same units as the resource. A binding constraint has zero slack at the solution. Nonbinding resources may have unused capacity.

5.2 Shadow prices

The shadow price of a resource is the marginal change in the optimal objective value caused by a one-unit increase in that resource’s availability, within an allowable range and holding other assumptions constant. For a profit-maximization problem, it is commonly measured in dollars per additional resource unit.

A shadow price is not automatically a market price. It is a model-based marginal value. If a resource has positive shadow price, acquiring a small additional amount may improve the objective, provided the model remains in the same sensitivity range. A nonbinding resource commonly has a zero shadow price, although degeneracy and alternative optima require care.

5.3 Scenario analysis

Scenario analysis changes a coherent set of assumptions, such as demand, prices, labor availability, or supply costs. A scenario table might include base, optimistic, and adverse cases. Unlike one-way sensitivity analysis, scenarios can represent relationships among variables—for example, high demand may also lead to higher material prices.

Robust decisions often perform reasonably well across several plausible scenarios rather than optimizing only one forecast. Scenario analysis is not the same as assigning probabilities; probabilities are needed when calculating expected values.

6. Simulation for Uncertainty and Risk

Optimization usually treats inputs as fixed unless uncertainty is explicitly modeled. Simulation represents uncertain inputs with probability distributions and repeatedly evaluates possible outcomes.

A basic Monte Carlo procedure is:

1. Specify uncertain inputs and their distributions.
2. Generate random values for one trial.
3. Calculate the resulting outcome.
4. Repeat for many trials.
5. Summarize the distribution of outcomes.
6. Test assumptions and validate the model.

If Y_k is the outcome in trial k, the simulated mean after N trials is:

The sample standard deviation is:

The mean and standard deviation describe simulated outcomes, not necessarily the true population values. More trials generally reduce simulation noise, but they cannot correct an inappropriate distribution or flawed model logic.

For an estimated mean, the standard error is approximately:

This formula assumes trials are independent or appropriately designed. Correlated trials require alternative analysis.

Example: Inventory risk

Suppose weekly demand is modeled as normally distributed with mean 1,000 units and standard deviation 150 units. A firm stocks 1,200 units. Under the normal model, the standardized value is:

The probability of stockout is P(D > 1,200), approximately 0.092 using the standard normal distribution. This estimate depends strongly on the normality assumption, and demand cannot realistically be negative. If demand is skewed or seasonal, empirical or alternative distributions may be more appropriate.

Simulation can incorporate reorder rules, lead-time uncertainty, lost sales, holding costs, and service penalties—features that may be difficult to capture in a simple formula.

Optimization and simulation together

A common approach is to optimize a decision using estimated inputs and then simulate the recommended policy under uncertain conditions. Another approach, simulation optimization, searches for decisions whose simulated performance is best. Because simulation outputs contain sampling error, comparisons should use sufficient replications, common random numbers where appropriate, and confidence intervals or statistical tests.

7. Decision Trees and Expected Monetary Value

Decision trees represent sequential decisions and uncertain events. Squares denote decisions, circles denote chance events, and terminal branches contain outcomes. They are especially useful when managers can observe information and then make later decisions.

The expected monetary value (EMV) of an action is:

where p_i is the probability of outcome i, V_i is its monetary value, and probabilities sum to 1. The EMV criterion is appropriate when probabilities are credible and the decision maker is risk-neutral, or when EMV is used as one input alongside risk analysis.

Worked Example: Capacity choice

A firm must choose a small or large facility. Demand may be high with probability 0.6 or low with probability 0.4. Payoffs are:

The small facility has:

The large facility has:

Under EMV, the large facility is preferred. However, the large facility has a much lower low-demand payoff. A risk-averse manager may prefer the small facility, especially if losses threaten liquidity or employment. Expected value should therefore be accompanied by downside measures, such as worst-case payoff, probability of loss, or expected shortfall.

The expected value of perfect information (EVPI) is:

In this example, perfect information would produce $0.6(160,000) + $0.4(70,000) = $124,000, so EVPI is $124,000 − $104,000 = $20,000. No information system should cost more than its decision value under the assumptions used.

8. Applied Case: Regional Food Distributor

Background

The following is an illustrative case. A regional food distributor supplies supermarkets from one warehouse. It must decide how many pallets of three products to ship during the next week. Product demand is uncertain, refrigerated capacity is limited, and shortages damage customer relationships.

Problem

The planner must balance contribution margin against storage and transport limits. Forecasts provide expected demand, but actual demand may be 20% higher or lower. The company also wants to examine whether additional refrigerated capacity is valuable.

Analysis and approach

Let x_j denote pallets shipped of product j. The first model maximizes expected contribution:

The model is solved for the base forecast. The team then performs three analyses:

1. Sensitivity: vary refrigerated capacity by ±10% and inspect the objective and product mix.
2. Scenarios: evaluate low, base, and high demand, with transport costs adjusted consistently.
3. Simulation: draw demand from calibrated distributions and calculate profit after shortage penalties and leftover spoilage.

The simulation should preserve dependencies. For example, high demand for related products may occur together; independently sampling every product could understate aggregate risk. The company can compare the base optimization policy with a policy that reserves capacity for high-margin products.

Outcome

Suppose the base LP uses all refrigerated space, while labor has slack. The positive shadow price on refrigerated space indicates that capacity is the immediate bottleneck. However, simulation may show that the base policy performs poorly in high-demand trials because it creates shortages, while a slightly lower expected-profit policy produces more stable service. Management may therefore purchase temporary refrigerated capacity only during high-demand scenarios, subject to supplier reliability and cost.

Learning points

The case illustrates that optimization identifies the best decision under stated inputs, while simulation reveals the distribution of results when inputs vary. Sensitivity analysis identifies which assumptions deserve managerial attention. The final decision should consider service commitments, food waste, supplier contracts, and operational feasibility—not only the nominal objective value.

9. Implementation, Governance, and Ethics

Prescriptive analytics influences resource allocation and can affect employees, customers, suppliers, and communities. Important controls include:

• Data governance: document sources, definitions, update dates, and ownership.
• Model governance: maintain version control, assumptions, validation evidence, and approval records.
• Auditability: retain inputs, solver settings, outputs, and overrides.
• Fairness: test whether constraints or recommendations disadvantage protected groups or vulnerable customers.
• Privacy: minimize personal data and apply access controls.
• Human oversight: define when managers may override recommendations and require reasons for material overrides.
• Stress testing: examine extreme but plausible conditions.
• Monitoring: compare predicted benefits and actual outcomes, then recalibrate.

AI-enabled optimization systems can automate scenario generation, forecasting, and natural-language explanations. They do not eliminate the need for valid objectives, transparent constraints, bias testing, or human accountability. An opaque recommendation can be operationally convenient but difficult to defend to regulators, employees, or customers.

10. Best Practices and Limitations

Good practice begins with a precise decision statement and a clear unit of analysis. Analysts should distinguish controllable decisions from forecasts and should avoid optimizing a metric that is merely easy to measure. Every constraint should have a business interpretation, and every objective coefficient should have a documented source.

Important limitations include:

• Incorrect forecasts can produce confidently wrong prescriptions.
• Linear relationships may fail at high volumes or near capacity.
• Correlated risks may be missed when inputs are sampled independently.
• Solver optimality does not guarantee implementation feasibility.
• Expected value can conceal severe downside risk.
• Historical data may encode inequitable practices.
• Model complexity can reduce explainability and adoption.

A strong professional report should state the recommendation, assumptions, binding constraints, sensitivity results, major risks, and implementation actions. It should also identify what evidence would cause the recommendation to change.

Chapter Summary

Prescriptive analytics converts business objectives and operational restrictions into recommended actions. Decision variables describe choices, objective functions measure value, and constraints define feasible solutions. Linear programming is useful when relationships are proportional and additive; integer and mixed-integer models represent indivisible and yes/no decisions. Feasibility, binding constraints, slack, and shadow prices help interpret optimization results.

Sensitivity analysis examines the effect of parameter changes, while scenario analysis evaluates coherent alternative futures. Monte Carlo simulation represents uncertainty and estimates the distribution of outcomes, but its validity depends on appropriate distributions, dependencies, and sufficient replications. Decision trees model sequential choices and uncertain events; expected monetary value is useful under credible probabilities but should be supplemented by risk analysis. In practice, optimization, simulation, and managerial judgment are complementary. Ethical governance, auditability, privacy, fairness, and monitoring are essential whenever recommendations affect people or scarce resources.

Key Terms

Prescriptive analytics: Analytics that recommends actions under objectives and constraints.

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

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

Objective function: A mathematical expression measuring the value to maximize or minimize.

Constraint: A condition limiting or requiring feasible decisions.

Feasible solution: A solution satisfying every model constraint.

Linear programming: Optimization with linear objectives and linear constraints.

Integer programming: Optimization requiring selected variables to be integers.

Binary variable: A variable restricted to 0 or 1.

Mixed-integer linear programming: A model containing continuous and integer variables.

Slack: Unused amount in a resource constraint.

Binding constraint: A constraint satisfied exactly at a solution.

Shadow price: Marginal objective improvement from one additional unit of a constrained resource.

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

Scenario analysis: Evaluation of decisions under coherent alternative assumptions.

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

Decision tree: A representation of sequential decisions and chance events.

Expected monetary value: Probability-weighted average payoff.

Expected value of perfect information: Value of knowing uncertain outcomes before deciding.

Model governance: Controls for documenting, validating, approving, monitoring, and auditing models.

Review Questions

Short Answer Questions

1. What are the roles of decision variables, objective functions, and constraints in a prescriptive model?
2. Distinguish a binding constraint from a constraint with slack.
3. Why may rounding an LP solution fail in an integer decision problem?
4. What does a positive shadow price indicate?
5. State two limitations of Monte Carlo simulation.

Long Answer Questions

1. Explain the assumptions of linear programming and discuss situations in which each assumption may be unrealistic.
2. Compare sensitivity analysis, scenario analysis, and simulation as methods for dealing with uncertainty.
3. Explain how decision trees represent sequential decisions and why expected monetary value may not match a risk-averse manager’s preference.
4. Describe how optimization and simulation can be combined in a business planning process.
5. Discuss governance and ethical controls required for an AI-enabled prescriptive analytics system.

Analytical / Application Questions

1. Formulate an LP for a firm producing two products with limits on labor, materials, and demand. Identify the units of every coefficient and explain how you would interpret a binding labor constraint.
2. A project has three possible outcomes with probabilities 0.2, 0.5, and 0.3 and payoffs of −$40,000, $30,000, and $90,000. Calculate its EMV and explain one reason a manager might reject it despite a positive EMV.
3. Design a simulation study for a retailer choosing an order quantity when demand, lead time, and spoilage are uncertain. Specify the inputs, output measures, validation checks, and decision rule.

Multiple Choice Questions with Answers

Question 1

Which component of a prescriptive model represents a controllable choice?

A. Decision variable

B. Parameter

C. Residual

D. Scenario probability

Correct Answer: A. Decision variable

Question 2

What does a binding resource constraint have at the solution?

A. Negative slack

B. Zero slack

C. Maximum slack

D. An undefined shadow price

Correct Answer: B. Zero slack

Question 3

In a maximization LP, a positive shadow price generally indicates that an additional unit of the resource would, within its valid range:

A. Reduce the objective value

B. Leave the objective unchanged in all cases

C. Improve the optimal objective value

D. Make every variable integer

Correct Answer: C. Improve the optimal objective value

Question 4

Which requirement is appropriate for a binary decision variable?

A. x ≥ 0

B. x ≤ 100

C. x ∈ {0,1}

D. x = 2.5

Correct Answer: C. x ∈ {0,1}

Question 5

A simulation produces 100 trials with outcomes whose sample mean is $50 and sample standard deviation is $20. The approximate standard error of the mean is:

A. $0.20

B. $2

C. $20

D. $200

Correct Answer: B. $2

Question 6

A decision tree is particularly useful when:

A. There are no uncertain outcomes

B. Decisions occur sequentially as information becomes available

C. All variables must be continuous

D. The objective cannot be quantified

Correct Answer: B. Decisions occur sequentially as information becomes available

Question 7

If an action has outcomes of $10,000 and $4,000 with probabilities 0.7 and 0.3, respectively, its EMV is:

A. $5,800

B. $7,000

C. $8,200

D. $10,000

Correct Answer: C. $8,200

Question 8

Which statement best distinguishes scenario analysis from one-way sensitivity analysis?

A. Scenario analysis changes only one coefficient

B. Scenario analysis changes a coherent set of assumptions

C. Scenario analysis never uses forecasts

D. Scenario analysis always assigns objective probabilities

Correct Answer: B. Scenario analysis changes a coherent set of assumptions

Question 9

Why can independent sampling of related product demands be problematic in simulation?

A. It may ignore correlation and understate aggregate risk

B. It guarantees infeasibility

C. It removes all sampling variation

D. It converts continuous variables into binary variables

Correct Answer: A. It may ignore correlation and understate aggregate risk

Question 10

Which practice most directly supports auditability of a prescriptive model?

A. Removing model assumptions from reports

B. Retaining inputs, assumptions, solver settings, and outputs

C. Using only the most complex algorithm

D. Allowing undocumented manual overrides

Correct Answer: B. Retaining inputs, assumptions, solver settings, and outputs