Prescriptive Analytics, Optimization and Simulation

Introduction

Business analytics becomes operationally valuable when it moves beyond describing what happened and predicting what may happen to recommending what an organization should do. Prescriptive analytics links forecasts, managerial objectives, constraints, uncertainty and available actions. It supports decisions such as how much inventory to hold, which suppliers to use, how to allocate capacity, when to schedule production, and how to route shipments.

A useful decision logic is:

Data → forecast → alternatives → model → decision → implementation → monitoring.

Forecasts are inputs, not decisions. A demand forecast may indicate that sales will increase, but it does not determine whether the firm should add overtime, subcontract production, build inventory, or accept some lost sales. Prescriptive analytics evaluates these alternatives against costs, service requirements, resource limitations and risk.

This chapter develops five connected tools: linear programming, integer and mixed-integer optimization, decision analysis, Monte Carlo simulation, and sensitivity and scenario analysis. Their appropriate use depends on the structure of the decision problem and the quality of the underlying data.

Learning Objectives

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

1. Distinguish descriptive, predictive and prescriptive analytics.
2. Formulate capacity, inventory, sourcing and logistics problems as optimization models.
3. Explain the role of integer variables, objective functions and constraints.
4. Apply decision trees and expected monetary value to decisions under uncertainty.
5. Design and interpret Monte Carlo simulation models.
6. Use sensitivity analysis, shadow prices and scenario planning responsibly.
7. Evaluate model assumptions, implementation risks, governance requirements and managerial implications.

From Forecasts to Decisions

A prescriptive model combines four elements:

• Decision variables: controllable choices, such as production quantity or supplier allocation.
• Objective: a measurable goal, such as minimizing total cost or maximizing contribution.
• Constraints: limits on capacity, cash, labor, supplier availability, policy or service.
• Parameters: forecasts, costs, lead times, probabilities and technical coefficients.

A deterministic optimization model assumes that parameters are known. A stochastic model represents uncertainty explicitly, often through probability distributions or scenarios. In practice, the forecast is often updated while the optimization is rerun, creating a rolling decision process.

The distinction between a feasible and an optimal solution is important. A feasible solution satisfies all constraints. An optimal solution is the feasible solution with the best objective value according to the model. Optimization cannot compensate for an incomplete objective, inaccurate data or an unrealistic constraint.

Linear Programming

Linear programming (LP) is appropriate when decision variables are continuous and both the objective and constraints are linear. Its general minimization form is:

subject to:

Here, x_j is decision variable j, c_j is its unit cost or negative contribution, a_ij is the amount of resource i used by one unit of activity j, and b_i is the available amount or required level of resource i. Units must be consistent: if c_j is dollars per unit, c_jx_j is dollars.

The linearity assumption means proportionality and additivity. Producing twice as much uses twice the labor and incurs twice the modeled cost; the total effect is the sum of individual effects. Discounts, setup costs, congestion and nonlinear yield violate simple linearity unless represented through additional variables or alternative models.

Product-mix example

A manufacturer produces products A and B. Product A contributes $40 per unit and uses 2 labor hours; B contributes $30 and uses 1 labor hour. Available labor is 100 hours and demand limits are 30 units of A and 60 units of B. Let x_A and x_B denote production quantities.

subject to:

Because A contributes $20 per labor hour and B contributes $30 per labor hour, the model first favors B. Producing 60 units of B uses 60 hours, leaving 40 hours for A, or 20 units. The resulting contribution is:

This is feasible and optimal for the stated model. If B demand were unrestricted, all labor would be assigned to B. The demand constraint for B therefore has economic significance.

Dual values and managerial interpretation

An LP solution can provide a shadow price, also called a dual value. It estimates the change in the optimal objective value from a one-unit relaxation of a binding constraint, within the range for which the current solution structure remains valid. For example, a shadow price of $12 per additional labor hour means that one additional hour could increase contribution by approximately $12, provided other assumptions remain unchanged. It is not automatically the market price of labor and should not be extrapolated indefinitely.

Integer and Mixed-Integer Optimization

Many operational decisions cannot be fractional. A firm may open 3 warehouses, not 3.4; a truck may be assigned to a route or not; a production line may incur a fixed setup cost. Integer programming imposes:

Binary variables use y_k ∈ {0,1}. A fixed-charge decision can be represented by:

where x is activity level, y indicates whether the activity is selected, and M is a valid upper bound. If y = 0, then x = 0; if y = 1, activity may occur up to M. M should be as tight as possible. An excessively large M can cause numerical instability and slow computation.

Mixed-integer linear programming (MILP) combines continuous and integer variables. It is widely used for facility location, vehicle assignment, workforce scheduling, lot sizing and supplier selection. Common solution methods include branch-and-bound, branch-and-cut and commercial or open-source optimization solvers. The solver's result must still be checked for feasibility, data errors and business interpretation.

Inventory and sourcing formulation

Suppose demand for a product is forecast for periods t = 1, …, T. Let x_t be regular production, o_t overtime production, [[MATH.I_t]] ending inventory, and s_t shortage or backorders. A basic inventory balance is:

where d_t is demand and initial inventory is known. If shortages are not permitted, set s_t = 0. A cost-minimization objective might be:

where c_x, c_o, h and p are regular-production, overtime, holding and shortage costs per unit. Capacity constraints can be written as x_t ≤ C_t and o_t ≤ O_t. The model assumes costs and demand parameters are known for the planning run; uncertainty can be addressed through scenarios or simulation.

Network and Logistics Models

A transportation model allocates shipments from sources i to destinations j. Let [[MATH.x_ij]] be units shipped, [[MATH.c_ij]] unit transportation cost, [[MATH.s_i]] supply and [[MATH.d_j]] demand.

subject to:

If total supply equals total demand and all must be used, equalities may replace the inequalities. The model can be extended with lane capacities, emissions, delivery-time constraints, fixed route costs and multiple products. A minimum-cost solution may not be preferred if it creates unacceptable disruption risk or violates strategic sourcing requirements.

Scheduling and Resource Allocation

Scheduling problems assign jobs to machines, workers or time periods. A basic assignment variable [[MATH.x_ij]] equals 1 if job i is assigned to machine or period j. Constraints ensure each job is assigned once and capacity is not exceeded. More advanced models represent precedence relationships, sequence-dependent setup times and changeovers.

The choice of objective matters. Minimizing total completion time differs from minimizing maximum lateness, labor cost or weighted tardiness. A mathematically optimal schedule can be operationally poor if it ignores maintenance windows, fatigue, skills, labor agreements or schedule stability.

Decision Analysis Under Uncertainty

Optimization is often complemented by decision analysis. A decision tree represents decision alternatives, uncertain events and consequences. If alternative a has outcomes k with probabilities p_k and monetary values V_ak, its expected monetary value (EMV) is:

with ∑ₖpₖ = 1. The preferred alternative under risk neutrality is the one with the highest EMV for gains or lowest expected cost for costs. EMV does not imply that outcomes will average out in one trial; it is a long-run expectation.

Worked decision example

A firm may launch a product or remain with its current product. A launch costs $100,000. If demand is high, the launch produces $260,000 revenue; if demand is low, revenue is $80,000. The probability of high demand is 0.6.

Launch outcomes after the launch cost are $160,000 and −$20,000. Thus:

If the current product generates $50,000 with certainty, launch has the higher EMV. However, risk-averse management may reject launch because of the $20,000 downside. The expected value of perfect information (EVPI) estimates the maximum amount worth paying to know the future state before deciding:

In this example, with perfect information the firm launches in the high-demand state and stays with the current product in the low-demand state. The expected value is 0.6($160,000) + 0.4($50,000) = $116,000, so EVPI is $28,000. This is an upper bound, not a guaranteed value of a forecast study.

Monte Carlo Simulation

Simulation evaluates system performance by repeatedly sampling uncertain inputs and observing outputs. It is especially useful when uncertainty, nonlinear relationships, queues or complex rules make analytical optimization difficult.

A Monte Carlo model normally follows these steps:

1. Define the decision question and output measures.
2. Identify uncertain inputs and collect evidence.
3. Select distributions and parameter values.
4. Generate random draws using a documented seed or random-number method.
5. Apply system logic for each trial.
6. Summarize the output distribution, not only its mean.
7. Validate the model and use it to compare decisions.

For example, annual demand might be modeled as normally distributed, but a nonnegative and skewed demand variable may be better represented by a lognormal or empirical distribution. Distribution choice should reflect historical data, business knowledge and forecast-error analysis.

If output values are Y₁, Y₂, …, Y_N, the simulated mean is:

The sample standard deviation is:

The standard error of the simulated mean is approximately s/√N when trials are independent. Increasing N reduces sampling error, but it does not correct a biased model or incorrect distribution.

Inventory simulation illustration

Assume daily demand is generated from an empirical distribution and replenishment arrives after a random lead time. For each simulated day, the model records opening inventory, demand, receipts, ending inventory, shortage and ordering decisions. After many replications, management can compare policies such as a reorder point of 40 units versus 60 units.

Relevant outputs include average inventory, probability of stockout, fill rate, shortage quantity and total cost. A policy with the lowest average cost may be unacceptable if its stockout probability exceeds the service target. Simulation therefore exposes trade-offs rather than producing a universally best answer.

For time-dependent systems, warm-up periods may be needed to remove artificial starting conditions. Replications with different random streams can support confidence intervals. Correlated inputs, such as demand and emergency freight cost, should not be sampled independently without justification.

Sensitivity Analysis and Scenario Planning

Sensitivity analysis changes one or more model inputs to determine how results respond. It can identify critical assumptions, break-even points and robust decisions. A simple percentage sensitivity measure is:

This ratio is dimensionless and is useful for comparing variables measured in different units. One-at-a-time analysis is easy to explain but may miss interactions. A structured design, such as a scenario matrix or global sensitivity analysis, is more appropriate when many inputs vary simultaneously.

Scenario planning constructs coherent combinations of assumptions, such as high demand with constrained supply, or low demand with abundant capacity. Scenarios are not predictions; they are plausible environments used to test strategy. A good scenario includes internally consistent demand, price, lead-time, labor and supply assumptions.

A robust decision performs acceptably across several plausible scenarios. A fragile decision performs extremely well in one scenario but poorly elsewhere. Scenario analysis should distinguish:

• Best/base/worst cases: often simple managerial communication.
• Stress tests: deliberately severe but plausible conditions.
• Probabilistic simulation: assigns distributions or probabilities.
• What-if analysis: changes selected inputs to examine consequences.

Illustrative Case: Northstar Medical Devices

Background

Northstar Medical Devices assembles a diagnostic unit using two suppliers. Supplier A is inexpensive but has variable lead time; Supplier B is more expensive but reliable. Monthly demand is forecast at 10,000 units, with forecast error and occasional demand surges. The assembly line has regular capacity of 9,000 units and overtime capacity of 2,000 units. Holding inventory is costly because components can become obsolete.

Problem

Management must decide monthly supplier quantities, overtime use and safety inventory while maintaining a target probability of fulfilling demand. The lowest unit purchase cost alone is not an adequate objective because late deliveries cause expedited freight and lost customer trust.

Analysis and approach

The team first creates a MILP for supplier allocation and production planning. Binary variables represent whether a supplier is activated; continuous variables represent purchase, production and inventory. Constraints include supplier capacities, minimum contractual quantities, assembly capacity and inventory balance. The objective includes purchase, production, holding, expedite and shortage penalties.

The model is then evaluated under demand and lead-time scenarios. A Monte Carlo simulation samples forecast errors and supplier lead times and applies the proposed policy. Outputs include fill rate, total cost, stockout probability and overtime usage.

Sensitivity analysis shows that the decision is most sensitive to Supplier A's late-delivery probability and the shortage penalty. A small change in purchase price has less impact than a change in lead-time reliability. The team therefore considers a dual-sourcing policy rather than concentrating all volume with Supplier A.

Outcome

The recommended policy costs slightly more in the expected case but keeps simulated stockout probability below the service threshold across the principal scenarios. It also reduces dependence on one supplier and limits emergency procurement. The recommendation is approved subject to monthly re-estimation of demand error and supplier performance.

Learning points

1. The cheapest nominal input may not minimize total system cost.
2. Optimization selects a policy under stated assumptions; simulation tests behavior under uncertainty.
3. A service constraint makes customer consequences explicit.
4. Robustness across scenarios can be more valuable than maximum performance in one forecast.
5. Governance requires monitoring and model revision after implementation.

Implementation, Governance and Technology

Successful prescriptive analytics requires more than a solver. Data must have clear ownership, definitions and timestamps. Forecast versions, cost assumptions, constraints and solver settings should be recorded so that decisions are reproducible. Model outputs should be reviewed by operational experts before execution.

AI and machine learning can improve demand forecasts, detect anomalies and estimate lead-time distributions. They do not remove the need to define objectives and constraints. An opaque recommendation may be unsuitable when managers must explain decisions to customers, employees, regulators or auditors. A practical architecture separates forecast generation, optimization, simulation, approval and execution while preserving traceability.

Important governance questions include:

• Is the objective aligned with organizational and stakeholder interests?
• Are service, safety, labor and environmental constraints represented?
• Does the model create discriminatory or geographically unfair allocation?
• Can a human override the recommendation with a documented reason?
• Are privacy, cybersecurity and commercial confidentiality protected?
• Are automated decisions monitored for drift and unintended consequences?

Challenges and Limitations

Prescriptive models face several recurring limitations:

• Parameter uncertainty: Forecasts, costs and probabilities are estimates.
• Model risk: A mathematically correct model can represent the wrong system.
• Data quality: Missing, delayed or inconsistent data can distort recommendations.
• Complexity: Large MILPs may require time limits and produce bounds rather than proven optima.
• Behavioral response: Suppliers, employees and customers may react to a policy.
• Objective conflict: Cost, resilience, sustainability, speed and fairness may conflict.
• False precision: A result such as $1,247,382 may imply more certainty than the inputs justify.
• Implementation gap: A feasible mathematical solution may be impossible to execute because of informal practices or unavailable information.

Best practice is to use models as decision support, combine quantitative evidence with domain expertise, test extreme conditions, communicate uncertainty and monitor realized outcomes.

Chapter Summary

Prescriptive analytics translates forecasts into recommended actions. Linear programming models continuous allocation problems using linear objectives and constraints. Integer and mixed-integer models represent indivisible choices, fixed charges and logical conditions. Inventory, sourcing, transportation and scheduling decisions can be formulated through variables, balances, capacities and service requirements.

Decision analysis uses probabilities and expected monetary value to compare alternatives under uncertainty, while Monte Carlo simulation estimates distributions of outcomes when system logic is complex. Sensitivity analysis identifies influential assumptions; scenario planning tests coherent alternative futures and supports robust strategy. Effective implementation requires valid data, transparent assumptions, governance, human review and continuous monitoring. Optimization is powerful, but its conclusions are only as credible as its objectives, constraints and uncertainty representation.

Key Terms

Prescriptive analytics: Analytics that recommends actions using objectives, constraints and evidence.

Decision variable: A controllable quantity selected by a model.

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

Constraint: A mathematical representation of a limitation or requirement.

Feasible solution: A solution satisfying every model constraint.

Linear programming: Optimization with continuous variables and linear relationships.

Integer programming: Optimization requiring some or all variables to be integers.

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

Binary variable: A variable restricted to 0 or 1.

Shadow price: Marginal objective improvement from relaxing a binding constraint within a valid range.

Transportation model: Network model allocating supply to destinations at minimum cost.

Inventory balance: Equation linking opening inventory, flows, demand and ending inventory.

Expected monetary value: Probability-weighted average of outcomes.

Expected value of perfect information: Maximum value of knowing the uncertain state before deciding.

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

Sensitivity analysis: Examination of how outputs change when inputs change.

Scenario planning: Evaluation of decisions under coherent alternative future conditions.

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

Model governance: Processes controlling data, assumptions, validation, use and monitoring of models.

Review Questions

Short Answer Questions

1. Distinguish predictive analytics from prescriptive analytics.
2. What are the four principal elements of an optimization model?
3. Why are binary variables useful in operational decisions?
4. What does a shadow price indicate?
5. Why should simulation outputs be examined as distributions rather than only as averages?

Long Answer Questions

1. Explain how a demand forecast can be integrated into an inventory optimization model. State the principal assumptions and limitations.
2. Compare linear programming, mixed-integer programming and Monte Carlo simulation in terms of purpose, strengths and weaknesses.
3. Explain expected monetary value and expected value of perfect information using an operational example.
4. Discuss how sensitivity analysis and scenario planning support capacity and sourcing decisions.
5. Describe a governance framework for deploying an automated prescriptive analytics recommendation.

Analytical / Application Questions

1. A firm has two plants, three markets and different production and transportation costs. Formulate a transportation and production LP, identify the variables and constraints, and explain what additional conditions would require integer variables.
2. Design a Monte Carlo simulation for a reorder-point inventory policy with uncertain demand and lead time. Specify inputs, outputs, validation checks and a suitable comparison between two policies.
3. A supplier offers a 5% price reduction but has a materially higher late-delivery probability. Develop a decision framework combining optimization, scenario analysis and service-risk measures. Explain how you would make the final recommendation.

Multiple Choice Questions with Answers

Question 1

Which element of an optimization model represents the quantity that management can control?

A. Decision variable

B. Shadow price

C. Random seed

D. Scenario probability

Correct Answer: A. Decision variable

Question 2

In a linear programming model, what does the constraint coefficient aᵢⱼ represent?

A. The total profit of the model

B. The amount of resource i used by one unit of activity j

C. The probability of scenario j

D. The number of decision variables

Correct Answer: B. The amount of resource i used by one unit of activity j

Question 3

A binary variable is restricted to which values?

A. −1 and 1

B. Any nonnegative real number

C. 0 and 1

D. Any integer greater than 1

Correct Answer: C. 0 and 1

Question 4

If a product contributes $40 and uses 2 labor hours, its contribution per labor hour is:

A. $10

B. $20

C. $38

D. $80

Correct Answer: B. $20

Question 5

What is the main purpose of an inventory-balance equation?

A. To calculate a supplier’s market share

B. To link opening inventory, flows, demand and ending inventory

C. To estimate a forecast confidence interval

D. To eliminate all uncertainty

Correct Answer: B. To link opening inventory, flows, demand and ending inventory

Question 6

If an alternative has outcomes of $100 and $20 with probabilities 0.7 and 0.3, its EMV is:

A. $44

B. $70

C. $76

D. $120

Correct Answer: C. $76

Question 7

Monte Carlo simulation is primarily used to:

A. Replace all managerial judgment

B. Evaluate outcomes under repeated random realizations of uncertain inputs

C. Guarantee an optimal solution

D. Convert every variable into an integer

Correct Answer: B. Evaluate outcomes under repeated random realizations of uncertain inputs

Question 8

A shadow price is best interpreted as:

A. The historical purchase price of a resource

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

C. The probability of a stockout

D. The fixed cost of opening a facility

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

Question 9

Which statement best describes scenario planning?

A. It predicts one certain future

B. It creates coherent alternative conditions to test decisions

C. It ignores interactions among assumptions

D. It is identical to calculating a sample mean

Correct Answer: B. It creates coherent alternative conditions to test decisions

Question 10

Why can a mathematically optimal solution still be unsuitable for implementation?

A. Optimization never uses constraints

B. The model may omit operational realities or use inaccurate assumptions

C. All optimal solutions contain fractional variables

D. Simulation always produces lower costs

Correct Answer: B. The model may omit operational realities or use inaccurate assumptions