Mathematical Modeling and Model Governance

Introduction

Operations research converts managerial concerns into structured decision problems. A manager may ask, “How much should we produce?”, “Which customers should receive scarce capacity?”, or “Where should a distribution centre be located?” These questions cannot be answered responsibly until the system, decisions, objectives, constraints, data and uncertainty have been made explicit.

A mathematical model is a purposeful abstraction of a real system. It does not reproduce every feature of reality. Instead, it retains features that are relevant to a decision and represents them using variables, parameters, relationships and assumptions. Model governance is the discipline of ensuring that models are appropriately designed, documented, tested, used, monitored and retired. It is therefore both a technical and managerial responsibility.

At postgraduate level, modeling requires more than writing an objective function. It requires judgment about boundaries, data quality, uncertainty, stakeholder interests, validity, ethical consequences and the conditions under which a model should not be used.

Learning Objectives

After studying this chapter, you should be able to:

1. Translate a managerial problem into decision variables, parameters, objectives and constraints.
2. Distinguish feasible solutions, optimal solutions and implementable decisions.
3. Classify deterministic, stochastic, static, dynamic, linear, nonlinear, discrete and simulation models.
4. Formulate and interpret a small operations-research model.
5. Evaluate data quality, assumptions, sensitivity, validity and limitations.
6. Design a proportionate governance process for operational models.
7. Identify risks involving bias, transparency, privacy, accountability and model misuse.

1. From managerial problem to mathematical model

A modeling exercise begins with a decision statement, not with an algorithm. The analyst should clarify:

• Who is the decision maker?
• What decisions can actually be controlled?
• Over what time horizon?
• What resources, policies or regulations restrict decisions?
• What outcomes matter, and to whom?
• Which quantities are known, estimated or uncertain?
• What would count as a feasible and implementable solution?

A useful modeling chain is:

managerial situation → decision problem → conceptual model → mathematical model → computational solution → managerial action → monitoring and learning

The chain is iterative. If the solution is operationally unacceptable, the model may have omitted a constraint, represented an objective poorly, or relied on unreliable data.

1.1 System boundary and level of abstraction

The system boundary identifies what is inside the model and what is treated as external. For example, a production-planning model may include machines, labour, inventory and demand, while treating supplier reliability and exchange rates as exogenous inputs. A boundary that is too narrow can omit important causes; one that is too broad can make the model expensive, opaque and difficult to validate.

The appropriate level of detail depends on the decision. A strategic location model may use annual demand by region, whereas a warehouse scheduling model may require hourly arrivals and handling times. A model is not “more accurate” merely because it contains more variables. Detail is valuable only when it improves decisions sufficiently to justify its cost and complexity.

1.2 Components of a mathematical model

Decision variables must represent controllable choices. Forecast demand is generally a parameter or random variable, not a decision variable. A variable should also have a clear unit and domain. For example, x_i could represent pallets shipped from plant i per week, with x_i ≥ 0. If whole vehicles are required, x_i may need to be integer.

Parameters may be certain, estimated or scenario-dependent. Distinguishing these categories is essential because a highly precise-looking parameter can create false confidence when its empirical basis is weak.

2. Objectives, constraints and feasibility

An objective function expresses what the model seeks to improve. A generic single-objective optimization problem is:

Here, x is the vector of decision variables, θ is the vector of parameters, f is the objective function, g_i and h_k represent inequality and equality constraints, and X specifies domains such as non-negativity or integrality. The formulation applies broadly, although the method used to solve it depends on properties such as linearity and convexity.

A feasible solution satisfies every constraint. An optimal solution is a feasible solution with the best objective value according to the model. An optimal solution may still be infeasible in practice if it violates an unmodeled labour agreement, requires unavailable technology or cannot be implemented quickly enough.

2.1 Hard and soft constraints

A hard constraint cannot be violated, such as a legal safety limit. A soft constraint expresses a preference that may be violated at a penalty, such as a target service level. Soft constraints can be represented with deviation variables. If actual staffing is s and the target is T, under-staffing may be represented by u ≥ T − s with penalty p_u u in the objective. The penalty must have a defensible interpretation; an arbitrary large number can distort priorities and create numerical difficulties.

2.2 Multiple objectives

Operations decisions commonly balance cost, service, quality, resilience and environmental performance. Three approaches are common:

1. Weighted objective: combine normalized criteria using managerial weights.
2. Lexicographic priority: optimize the most important objective first, then optimize secondary objectives without worsening the primary result.
3. ε-constraint method: optimize one objective while imposing acceptable bounds on others.

Weights are not purely technical inputs. They encode values and should be agreed with relevant stakeholders. A weighted objective such as αC + βE is meaningful only when C and E are scaled comparably and α and β are justified.

3. A worked formulation: production planning

Consider a firm producing products A and B during one week. Each unit of A yields a contribution of $40 and requires 2 labour hours and 3 kg of material. Each unit of B yields $30 and requires 1 labour hour and 4 kg of material. Available resources are 100 labour hours and 180 kg of material. At most 40 units of B can be sold.

Let:

• x_A = units of A produced;
• x_B = units of B produced.

The deterministic linear programming model is:

The objective has units of dollars per week. Each constraint is unit-consistent: labour coefficients are hours per unit and the right-hand side is hours; material coefficients are kilograms per unit and the right-hand side is kilograms.

To inspect candidate corner points, consider the intersections of binding constraints. The labour and material constraints intersect where:

From the first equation, x_B = 100 − 2x_A. Substitution gives 3x_A + 4(100 − 2x_A) = 180, so x_A = 220/5 = 44 and x_B = 12. The contribution is 40(44) + 30(12) = 2,120 dollars. This point satisfies the market limit and is feasible.

Other relevant corner points include (0,40), yielding $1,200; (0,0), yielding $0; and the material-axis point (60,0), yielding $2,400. However, at (60,0), labour usage is 120 hours, exceeding the labour limit, so it is infeasible. The feasible intersection of the labour axis and the material constraint is not automatically feasible because all constraints must be checked. Evaluating the feasible vertices shows that (44,12) is optimal among the relevant feasible points only if no other feasible vertex gives more; in this case, the point where material binds with x_B = 40 is x_A = 20/3 ≈ 6.67, with contribution approximately $1,466.67, while the labour-axis feasible point (50,0) gives $2,000. Thus the best listed feasible vertex is (44,12), with contribution $2,120.

This example illustrates several modeling lessons. First, a market restriction can be inactive at the optimum. Second, resource utilization should be reported: at the solution, labour use is 2(44)+12=100 hours and material use is 3(44)+4(12)=180 kg. Third, the result depends on contribution margins, capacities, demand assumptions and divisibility. If products must be produced in whole units, an integer model is more appropriate, and (44,12) happens to remain integer but that need not generally occur.

4. Classification of operations-research models

4.1 Deterministic and stochastic models

A deterministic model treats all inputs as known constants for the modeled period. It is appropriate when variability is negligible, decisions are made using fixed scenarios, or a planning approximation is useful.

A stochastic model represents uncertain inputs with probability distributions, scenarios or stochastic processes. For example, demand may be represented by a random variable D. If profit under decision x is Π(x,D), the expected-profit objective is:

where E denotes expectation. Maximizing expected profit does not necessarily minimize downside risk. A risk-sensitive model may constrain the probability of loss or use a measure such as conditional value at risk. Expected-value models can also be misleading when decisions are difficult to reverse or when extreme outcomes matter disproportionately.

4.2 Static and dynamic models

A static model examines one period or assumes that time does not materially affect relationships. A dynamic model represents decisions and system states over time. A basic inventory balance is:

where I_t is ending inventory in period t, q_t is replenishment or production, and d_t is demand. The identity assumes that inventory is measured consistently and does not explicitly include spoilage, returns or lost sales. Those features require additional terms.

Dynamic models may include state variables, transition equations and intertemporal costs. The distinction is important because a decision that is attractive in one period may create future shortages, overtime or obsolescence.

4.3 Linear, nonlinear, discrete and simulation models

A linear model has objectives and constraints that are linear in the decision variables. Linear programming is computationally powerful and interpretable, but linearity may be inappropriate when costs, processing times or substitution effects change with volume.

A nonlinear model includes products, powers, nonlinear functions or nonlinear constraints. It may represent economies of scale, congestion or risk more realistically, but can be harder to solve and may have local optima.

A discrete or integer model requires some variables to take integer or binary values. It is necessary for choices such as opening a facility, assigning a vehicle or selecting a project. Integer restrictions usually increase computational difficulty.

A simulation model estimates system performance by reproducing the operation of a system under specified rules and random inputs. Simulation is especially useful when analytical equations are unavailable, but it generally evaluates policies rather than directly producing an optimum unless combined with optimization or search. Its outputs require multiple replications and statistical interpretation.

5. Data quality and model validity

A model can be mathematically correct yet decisionally wrong. Data quality should be assessed across several dimensions:

• Accuracy: Does a value represent the real quantity?
• Completeness: Are important records missing?
• Consistency: Do definitions agree across systems?
• Timeliness: Are data current for the decision horizon?
• Granularity: Are observations sufficiently detailed?
• Traceability: Can values be linked to their source and transformation?
• Representativeness: Do historical data reflect the population and conditions of intended use?

Data preparation must document exclusions, imputation, aggregation, unit conversion and outlier treatment. A forecast based on sales data may be biased if historical stockouts caused observed sales to be lower than true demand.

5.1 Verification, validation and credibility

Verification asks whether the model was built correctly according to its specification: equations, code, units and logic. Validation asks whether the model is an adequate representation for its intended purpose. A model may pass verification but fail validation if it omits a critical operational mechanism.

Validation evidence can include:

• face validation by domain experts;
• comparison with historical outcomes;
• back-testing on data not used for calibration;
• extreme-condition tests;
• sensitivity and stress tests;
• comparison with simpler benchmark models;
• pilot implementation and monitoring.

No single test proves that a model is universally valid. Validity is contextual: a model suitable for weekly capacity planning may be unsuitable for real-time dispatching.

5.2 Sensitivity and robustness

Sensitivity analysis changes one or more inputs to examine their effect on outputs. If a small change in demand reverses the recommended decision, management should treat the result cautiously. Scenario analysis changes coherent groups of assumptions, such as a recession scenario or supplier disruption scenario.

Robust decisions perform acceptably across plausible conditions, even if they are not optimal under one forecast. Robustness is particularly valuable when data are uncertain and corrective action is costly.

6. Model governance

Model governance is the framework of roles, controls and evidence used to manage model risk throughout the model life cycle. A practical life cycle is:

1. Problem definition: document purpose, users, decisions and boundaries.
2. Design: specify variables, objectives, constraints, data and assumptions.
3. Development: implement equations, code and data pipelines.
4. Independent review: test logic, data, performance and limitations.
5. Approval: authorize use for a defined purpose and population.
6. Deployment: control versions, access, interfaces and overrides.
7. Monitoring: track accuracy, drift, constraint violations and outcomes.
8. Change management: assess modifications before release.
9. Retirement: archive evidence and withdraw obsolete models.

A model inventory should record the owner, purpose, version, materiality, data sources, validation status, approval date, limitations and next review date. Governance should be proportionate. A spreadsheet used for a low-impact internal estimate does not need the same controls as a model allocating emergency capacity or determining employee schedules.

6.1 Roles and accountability

The model owner is accountable for business purpose and continued suitability. The developer builds and documents the model. An independent validator challenges design, data and performance. The user applies outputs within authorized conditions. A governance committee may approve high-impact models and resolve conflicts between operational urgency and control requirements.

Separation of development and validation reduces self-confirmation risk. However, independence should not mean absence of operational knowledge; validation requires both technical competence and understanding of the business context.

6.2 Transparency and explainability

Transparency means that stakeholders can understand the model’s purpose, inputs, logic, limitations and decision rights. Explainability is not merely a technical description of an algorithm. A transparent operational model should answer:

• What decision does it support?
• Which inputs materially influence results?
• What assumptions are embedded?
• Under what conditions is it unreliable?
• Who can override it, and how is the override recorded?
• What recourse exists for affected parties?

Complexity should be justified by measurable decision value. An opaque model can be acceptable in limited contexts if its performance, limitations, controls and human accountability are strong, but opacity is not a substitute for governance.

6.3 Ethical, legal and operational risks

Models can reproduce historical inequity, allocate scarce services unfairly, expose personal information or encourage unsafe work intensification. Governance should therefore consider:

• purpose limitation and data minimization;
• privacy, access control and retention;
• fairness across relevant groups;
• safety and human override;
• accountability for adverse outcomes;
• compliance with applicable employment, consumer, sectoral and data-protection rules;
• cybersecurity and resilience against manipulation.

Fairness is context-dependent. Equal treatment, equal outcomes and priority to vulnerable groups are different principles. The organization must make the normative choice explicit rather than hiding it in data or code.

7. Illustrative case: hospital elective-surgery scheduling

This is a fictional illustrative case designed for teaching purposes.

Background and problem

A regional hospital must schedule elective surgeries over a two-week horizon. Operating rooms, surgeons, nurses and postoperative beds are limited. The hospital wants to reduce waiting time while protecting emergency capacity and complying with maximum working hours.

Analysis

An initial proposal maximized the number of completed surgeries using average procedure times. Review identified three weaknesses: procedure times were variable, emergency admissions displaced elective work, and the objective ignored waiting-time equity across clinical groups. Historical data also contained cancellations caused by staffing shortages, making observed completion rates a poor measure of unconstrained demand.

Approach and solution

The redesigned model used binary variables for assigning each patient to a time slot, capacity constraints for rooms and staff, a reserved emergency-capacity constraint, and penalties for postponement. Procedure duration was represented using conservative planning times, while scenario analysis tested high-demand and emergency-surge conditions. The model did not autonomously approve patients; a clinical scheduling team reviewed recommendations and could document exceptions.

Governance controls included a data dictionary, version-controlled code, independent validation, weekly monitoring of cancellations and waiting-time distribution, and a rule that the model could not reduce protected emergency capacity without executive approval.

Outcome and learning points

In pilot testing, the hospital obtained a more balanced schedule and clearer visibility of trade-offs. The principal benefit was not simply a higher utilization rate; it was improved coordination between clinical priorities, capacity limits and escalation procedures. The case demonstrates that a technically feasible schedule is insufficient. A credible model must represent uncertainty, respect policy constraints, expose value judgments and remain subject to accountable human review.

8. Technology and contemporary developments

Cloud computing, optimization solvers, process mining and machine learning can improve model development and monitoring. Process mining can reveal actual process variation before a queueing or scheduling model is specified. Machine learning can support demand forecasts, while optimization converts forecasts into decisions. These components should not be treated as interchangeable: forecast accuracy does not guarantee decision quality, and an optimized decision can still be harmful if the objective is incomplete.

Generative AI can assist with documentation, code suggestions and scenario generation, but generated content requires verification. It may fabricate data sources, omit constraints or produce code with subtle errors. Sensitive data should not be entered into unapproved systems. Human experts remain responsible for model specification, validation and use.

Contemporary practice increasingly emphasizes model cards or model factsheets, reproducible pipelines, data lineage, continuous monitoring and challenger models. These practices improve institutional memory and make it easier to detect model drift when operating conditions change.

Best-practice checklist

Before approving an operations-research model, ask:

1. Is the decision, user and time horizon explicit?
2. Are decision variables genuinely controllable and correctly typed?
3. Are units, signs and domains documented?
4. Are all material objectives and constraints represented?
5. Are assumptions defensible and visible?
6. Are data sources, transformations and uncertainties traceable?
7. Has the model been verified independently of its developer?
8. Has it been validated for the intended decision context?
9. Have sensitivity, stress and failure-mode tests been conducted?
10. Are accountability, privacy, fairness, override and monitoring arrangements clear?

Chapter Summary

Mathematical modeling translates operational situations into decision variables, parameters, objectives, constraints and assumptions. Feasibility precedes optimality, and mathematical optimality does not guarantee operational implementability. Models may be deterministic or stochastic, static or dynamic, linear or nonlinear, continuous or discrete, analytical or simulation-based. The choice should follow the decision context and the available evidence.

Reliable modeling requires attention to units, data lineage, uncertainty, verification, validation, sensitivity and robustness. Model governance extends across definition, development, independent review, approval, deployment, monitoring, change and retirement. It assigns accountability and addresses transparency, privacy, fairness, safety and human oversight. The central professional principle is that a model is a decision-support instrument, not an unquestionable authority.

Key Terms

Decision variable: A controllable quantity selected by the decision maker.

Parameter: An input treated as fixed or estimated within a model.

Objective function: A mathematical representation of the criterion to maximize or minimize.

Constraint: A mathematical requirement limiting allowable decisions.

Feasible solution: A decision satisfying all stated constraints and domains.

Optimal solution: The best feasible solution according to the specified objective.

Assumption: A simplification or proposition accepted for model construction.

Deterministic model: A model whose inputs are treated as known constants.

Stochastic model: A model that represents uncertainty using random variables, distributions or scenarios.

Static model: A model that does not explicitly represent change over time.

Dynamic model: A model containing time-dependent states, decisions or transitions.

Linear programming: Optimization with linear objective and constraint relationships.

Integer programming: Optimization requiring selected variables to take integer or binary values.

Simulation: Computational imitation of system behaviour under specified inputs and rules.

Verification: Checking whether the model was implemented according to its specification.

Validation: Assessing whether the model is adequate for its intended purpose.

Sensitivity analysis: Studying output changes caused by changes in inputs or assumptions.

Model risk: The possibility of harmful decisions arising from model error, misuse or inappropriate application.

Model governance: Policies, roles and controls for managing models throughout their life cycle.

Data lineage: The documented origin, transformation and movement of data.

Review Questions

5 Short Answer Questions

1. What is the difference between a decision variable and a parameter?
2. Define a feasible solution and explain why it may not be implementable.
3. Distinguish verification from validation.
4. Give two examples of data-quality dimensions relevant to operations models.
5. Why should model assumptions be documented?

5 Long Answer Questions

1. Explain the complete process of translating a managerial problem into a mathematical model.
2. Compare deterministic optimization, stochastic optimization and simulation, including appropriate applications and limitations.
3. Discuss how multiple objectives can be represented and why weights require managerial justification.
4. Explain the major stages and responsibilities in a model-governance life cycle.
5. Evaluate the ethical and operational risks of using opaque decision-support models.

3 Analytical / Application Questions

1. Formulate a production-planning model for three products with labour, material and demand constraints. Clearly define variables, units, objective and assumptions.
2. A warehouse model recommends staffing based on historical order volumes, but historical data exclude orders lost during stockouts. Diagnose the modeling risk and propose validation and data-improvement steps.
3. Design a governance plan for an AI-assisted hospital scheduling model. Specify approval, independent validation, monitoring, human override and escalation controls.

Multiple Choice Questions with Answers

Question 1

Which item is a decision variable in a production-planning model?

A. Available machine hours

B. Forecast demand

C. Units of product manufactured

D. Unit contribution margin

Correct Answer: C. Units of product manufactured

Question 2

A feasible solution is one that:

A. Has the highest possible objective value

B. Satisfies all stated constraints and variable domains

C. Uses every available resource completely

D. Has only integer parameter values

Correct Answer: B. Satisfies all stated constraints and variable domains

Question 3

In the production example, the contribution from producing 44 units of A and 12 units of B is:

A. $1,466.67

B. $2,000

C. $2,120

D. $2,400

Correct Answer: C. $2,120

Question 4

Which model explicitly represents uncertain inputs using distributions or scenarios?

A. Deterministic model

B. Stochastic model

C. Static model

D. Linear identity

Correct Answer: B. Stochastic model

Question 5

What does verification primarily assess?

A. Whether the model is adequate for every possible use

B. Whether the model was implemented according to its specification

C. Whether stakeholders agree with the objective weights

D. Whether the model guarantees fair outcomes

Correct Answer: B. Whether the model was implemented according to its specification

Question 6

Which is an example of a soft constraint?

A. A statutory safety limit that cannot be exceeded

B. A physical conservation equation

C. A preferred service target that may be violated with a penalty

D. A variable-domain requirement that a binary variable be 0 or 1

Correct Answer: C. A preferred service target that may be violated with a penalty

Question 7

Why is sensitivity analysis useful?

A. It proves that a model is universally valid

B. It examines how outputs change when inputs or assumptions change

C. It replaces the need for data validation

D. It ensures that all constraints are binding

Correct Answer: B. It examines how outputs change when inputs or assumptions change

Question 8

Which governance practice most directly supports reproducibility?

A. Removing model documentation after deployment

B. Using undocumented manual overrides

C. Version control and documented data lineage

D. Allowing each user to modify the objective function

Correct Answer: C. Version control and documented data lineage

Question 9

What is a major limitation of maximizing expected profit under uncertainty?

A. It cannot use any numerical data

B. It necessarily minimizes downside risk

C. It may ignore the importance of extreme or irreversible outcomes

D. It requires every variable to be binary

Correct Answer: C. It may ignore the importance of extreme or irreversible outcomes

Question 10

In the inventory balance I_t = I_t−1 + q_t − d_t, what does q_t represent?

A. Ending inventory

B. Demand in period t

C. Replenishment or production in period t

D. Previous-period inventory

Correct Answer: C. Replenishment or production in period t