Operations Analytics and Decision Models

Introduction and context

Operations management increasingly depends on the ability to convert process data into defensible decisions. Quality records, production schedules, customer orders, equipment sensors, labour data and supplier information can reveal where a process is unstable, what is likely to happen next, and which intervention is most valuable. Operations analytics provides the methods for doing this systematically.

Analytics is not simply the production of charts. It links data, models, assumptions and managerial action. A useful operational analysis should answer four questions:

1. What has happened or what is happening?
2. What is likely to happen?
3. What could be done?
4. Which action is preferable, given cost, risk, constraints and uncertainty?

These questions correspond broadly to descriptive, predictive and prescriptive analytics. Decision models extend analytics by representing alternatives, consequences, probabilities, constraints and preferences. In Total Quality Management, these methods support evidence-based improvement, prevention of defects, capacity planning, service-level management and continual improvement.

Learning objectives

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

• distinguish descriptive, diagnostic, predictive and prescriptive analytics;
• formulate operational problems using relevant measures and assumptions;
• construct and evaluate decision trees under risk and uncertainty;
• apply expected value, expected opportunity loss and value-of-information reasoning;
• formulate basic linear programming and interpret its solution;
• analyse single-server queues and explain the effects of utilisation;
• use simulation appropriately when analytical models are inadequate;
• perform sensitivity and scenario analysis;
• design operational dashboards that support action rather than merely reporting;
• evaluate ethical, governance and implementation issues in analytics.

1. The analytics hierarchy in operations

Descriptive and diagnostic analytics

Descriptive analytics summarises observed performance. Typical measures include throughput, cycle time, first-pass yield, defect rate, on-time delivery, inventory turnover and overall equipment effectiveness. Diagnostic analytics investigates why performance occurred. It may use stratification, Pareto analysis, control charts, regression, process mapping or cause-and-effect analysis.

A rate must always be defined carefully. If 18 defective units are found in a sample of 600 units, the observed defect proportion is:

Here, d is the number of defective units and n is the number inspected. The calculation assumes that the inspection method is consistent and that the sample is relevant to the population being described. A defect proportion is not automatically a process capability measure; capability requires specification limits and an appropriate model of process variation.

Useful descriptive measures include:

Diagnostic analysis must distinguish correlation from causation. For example, overtime may be associated with defects because both are caused by demand surges. Reducing overtime alone may not remove the underlying cause. Process knowledge, controlled comparisons and appropriate experimental designs are therefore important.

Predictive analytics

Predictive analytics estimates future outcomes from historical and current information. Examples include forecasting demand, predicting equipment failure, estimating late deliveries and identifying transactions likely to contain errors.

A simple time-series forecast may use a moving average:

Dₜ is demand in period t, k is the number of periods included, and Fₜ₊₁ is the forecast for the next period. This method is appropriate when recent observations are informative and strong trend or seasonality is absent. It reacts slowly to sudden structural changes and should not be used uncritically for seasonal demand.

Exponential smoothing gives greater weight to recent data:

where 0 ≤ α ≤ 1 is the smoothing constant. A larger α responds more rapidly to recent changes; a smaller α produces a more stable forecast. Forecast accuracy can be monitored with measures such as mean absolute deviation:

The units of MAD are the units of demand. Forecast error should be assessed on data not used to fit or tune the model. Accuracy alone is insufficient: a highly accurate model that cannot be explained, maintained or used in time may have little operational value.

Prescriptive analytics

Prescriptive analytics evaluates possible actions. It may combine forecasts with optimisation, simulation, rules or decision analysis. Examples include selecting a production mix, allocating staff, choosing preventive-maintenance intervals or deciding how much safety stock to hold.

A prescriptive model requires:

• decision variables: what the organisation can control;
• an objective: what is to be maximised or minimised;
• constraints: physical, financial, legal or policy limits;
• parameters: demand, cost, capacity and time estimates;
• uncertainty treatment: scenarios, probability distributions or robust limits.

The distinction between prediction and prescription is essential. A model may predict a high probability of late delivery, but it does not by itself determine whether to expedite, reallocate capacity, change the promise date or accept the risk.

2. Decision-making under certainty, risk and uncertainty

Under certainty, the consequences of each alternative are assumed known. Under risk, possible states of nature and their probabilities are estimated. Under uncertainty, probabilities are unavailable, disputed or not credible. The same decision can produce different recommendations depending on this classification.

Expected monetary value

Suppose an organisation chooses action a and the uncertain state is s. If Vₐₛ is the payoff and P(s) is the probability of state s, expected monetary value is:

The expected-value rule selects the action with the highest EMV when payoffs are monetary, probabilities are credible and the decision-maker is reasonably risk neutral. For cost minimisation, the action with the lowest expected cost is selected.

Decision trees

A decision tree represents decisions as branches from decision nodes and uncertain events as branches from chance nodes. Each terminal branch has a consequence. The tree is evaluated from right to left by calculating expected values at chance nodes and selecting the best alternative at decision nodes.

Consider a capacity decision with the following estimated payoffs:

For expansion:

For no expansion:

Expansion has the larger EMV, although it has a loss in the low-demand state. The decision should also consider cash-flow capacity, risk tolerance, strategic flexibility and whether the probabilities are reliable.

Expected opportunity loss and value of information

Opportunity loss is the difference between the best payoff available after the state is known and the payoff from the chosen action. Expected opportunity loss is:

The expected value of perfect information is the difference between the expected payoff with perfect information and the best EMV without it:

For the example, the best payoff under high demand is $120,000 and under low demand is $20,000. Thus:

Therefore EVPI = $80,000 − $60,000 = $20,000. No information-gathering activity should cost more than its expected value unless it provides additional strategic or ethical benefits not captured in the monetary model.

Decisions under uncertain probabilities

When probabilities are not credible, analysts may use scenario analysis, minimax, maximax, minimax regret, Laplace equal probabilities or a weighted optimism rule. These are not universally correct; they express different attitudes toward ambiguity. A robust decision often performs acceptably across several plausible scenarios rather than optimising one fragile forecast.

3. Optimisation and linear programming

Optimisation identifies the best feasible decision according to a stated objective. Linear programming (LP) applies when the objective and constraints are linear and decision variables may be continuous or, in some extensions, integer-valued.

A general LP is:

subject to:

and:

Here, xⱼ are decision variables, cⱼ are objective coefficients, aᵢⱼ are resource-consumption coefficients, and bᵢ are resource limits. Coefficients must use compatible units.

Worked production-mix example

A plant produces Standard (S) and Premium (P) units. Each Standard unit contributes $40 and requires 2 labour hours and 1 machine hour. Each Premium unit contributes $60 and requires 1 labour hour and 2 machine hours. Available capacity is 100 labour hours and 80 machine hours.

The model is:

subject to:

The intersection of the two binding constraints is found from 2S + P = 100 and S + 2P = 80. Solving gives S = 40 and P = 20. Contribution is therefore:

The corner solutions are (50, 0), contribution $2,000, and (0, 40), contribution $2,400. Thus (40, 20) is the best feasible corner for this model, with 100 labour hours and 80 machine hours fully used.

The result is not automatically an implementation instruction. Demand limits, setup times, minimum batch sizes, quality requirements, maintenance windows and integer constraints may make the basic LP unrealistic. If units cannot be fractional, integer programming is more appropriate. If relationships are nonlinear, nonlinear programming or simulation may be needed.

Shadow prices and sensitivity

A shadow price estimates the marginal change in the optimal objective value from relaxing a binding resource constraint by one unit, within the allowable range over which the current solution remains valid. It is a managerial value, not necessarily a market price. Sensitivity analysis asks how the solution changes when costs, capacities, demand or quality assumptions change.

Important cautions include:

• a shadow price is local, not valid for unlimited capacity increases;
• changing several coefficients simultaneously can invalidate a simple interpretation;
• an optimal solution may be operationally infeasible if the model omits constraints;
• an apparently precise optimum can conceal substantial parameter uncertainty.

4. Queuing models and service capacity

Queues arise when stochastic arrivals require service from limited capacity. They are central to hospitals, call centres, repair facilities, inspection stations, warehouses and production lines.

The basic single-server M/M/1 model assumes Poisson arrivals, exponential service times, one server, first-come-first-served discipline, infinite waiting space and a stable system. Let λ be the mean arrival rate and μ the mean service rate, both in jobs per unit time. Utilisation is:

The stability condition is ρ < 1. For an M/M/1 queue, the expected number in the system and expected time in the system are:

By Little’s Law:

The expected number waiting and waiting time are:

Units must be consistent. If λ and μ are measured per hour, W and Wq are in hours, while L and Lq are numbers of jobs.

Suppose arrivals average 8 customers per hour and one employee serves 10 per hour. Then ρ = 0.8 and:

The average time in the system is 30 minutes. The expected number in the system is L = 8 × 0.5 = 4 customers. As utilisation approaches 1, waiting time increases sharply. This explains why operating a service process at nominal full capacity can produce poor customer experience.

Real systems often violate M/M/1 assumptions. Multiple servers, finite capacity, priority rules, batch arrivals, non-exponential service times and abandonment require other queueing models or simulation. A queueing result is a model-based estimate, not a guaranteed observation.

5. Simulation for complex processes

Simulation imitates the operation of a system over time using specified rules and random inputs. It is useful when analytical solutions are unavailable or when interactions among variability, capacity, routing, failures and priorities are important.

A discrete-event simulation typically contains:

1. entities, such as jobs or customers;
2. resources, such as machines or staff;
3. events, such as arrival, completion or failure;
4. state variables, such as queue length and machine status;
5. probability distributions for uncertain inputs;
6. performance outputs, such as waiting time, throughput and defects.

A credible simulation study requires verification and validation. Verification asks whether the model was built correctly. Validation asks whether it represents the real system adequately for the intended purpose. Historical data, subject-matter review, extreme-condition tests and comparison with known cases support validation.

Simulation outputs are estimates. Replications with different random-number streams are needed to estimate sampling variation. A confidence interval for a mean output may be represented as:

where x̄ is the mean across r replications, s is the sample standard deviation of replication results, and t is the relevant t-distribution critical value. This expression assumes independent, approximately normally distributed replication means. Warm-up periods may be required for terminating systems that begin empty but represent an ongoing operation.

Simulation should not be used to disguise weak assumptions. A detailed model with inaccurate arrival, service or failure data can produce confidently wrong recommendations.

6. Sensitivity, scenario and risk analysis

Sensitivity analysis changes one or more assumptions to determine whether a conclusion is stable. A one-way analysis changes one input at a time; a two-way analysis examines interactions between two inputs; scenario analysis changes a coherent set of assumptions, such as a recession, demand surge or supplier disruption.

For an output y depending on input x, a local sensitivity can be approximated by:

The units are output units per input unit. A dimensionless elasticity is:

Break-even analysis identifies the input at which two alternatives have equal value. If fixed costs are F, selling price is p and variable cost per unit is v, the break-even volume is:

This applies when p > v, costs are linear and the relevant output range is appropriate. It does not incorporate uncertainty, capacity limits or multiple products unless extended carefully.

Risk analysis should distinguish probability from impact. A low-probability event with severe consequences may require controls even when its expected monetary value is modest. Risk registers, fault-tree analysis, failure mode and effects analysis, stress testing and contingency planning complement quantitative models.

7. Operational dashboards and decision support

A dashboard should support a defined decision audience. Senior managers may need trends in customer-critical quality, cost and delivery; supervisors may need current queue length, takt adherence and abnormal conditions; process engineers may need control-chart signals and parameter distributions.

A useful dashboard contains:

• a clear purpose and owner;
• a small set of operationally meaningful measures;
• definitions, units, time windows and data sources;
• targets and control limits distinguished from specifications;
• drill-down from aggregate results to process, product, shift or supplier;
• visual indication of trend, variation and exception;
• a documented action or escalation rule.

For example, an on-time delivery percentage without order-volume context can mislead. A rate of 95% may represent 19 of 20 orders, whereas 95,000 of 100,000 orders has different operational significance. Dashboards should avoid false precision, inappropriate traffic-light thresholds and excessive metrics that dilute attention.

Control charts are especially useful when distinguishing common-cause from special-cause variation. A dashboard should not convert every fluctuation into an emergency. Nor should it conceal a persistent shift behind a favourable monthly average.

8. Data, AI and governance

Modern operations analytics may use machine learning for anomaly detection, predictive maintenance, computer vision inspection and demand forecasting. Such systems can process complex patterns, but they require representative data, stable definitions and monitoring for model drift.

Governance requirements include:

• data lineage and version control;
• access control and protection of personal or commercially sensitive data;
• documented assumptions and model ownership;
• validation before deployment;
• monitoring of accuracy, bias, drift and false alarms;
• human review for consequential decisions;
• an audit trail explaining recommendations.

An AI system trained on historical inspection decisions may reproduce past inconsistency or discrimination. A predictive maintenance model may favour assets with abundant sensor data and neglect poorly instrumented equipment. Automation bias may cause staff to accept a recommendation without questioning it. Human expertise remains essential for unusual conditions, ethical judgement and accountability.

9. Illustrative case: reducing waiting time in a diagnostic laboratory

Background

A regional laboratory receives specimen requests from outpatient clinics. Complaints concern long turnaround times during the morning peak. The laboratory has one accessioning desk and two testing instruments. Management is considering either adding a clerk or changing appointment and batching rules.

Problem

The average arrival rate during the peak is 18 specimens per hour. The accessioning clerk can process 24 specimens per hour, but arrivals are uneven. A simple M/M/1 approximation gives ρ = 18/24 = 0.75 and:

This estimate excludes downstream testing queues and therefore cannot explain the full turnaround time.

Analysis

Process mapping showed that specimens arrived in batches when clinic transport vehicles reached the laboratory. Testing instruments were sometimes idle while accessioning was overloaded. Historical data also showed that urgent and routine specimens were mixed in one queue. A simulation was built with time-varying arrivals, two testing resources, urgent priority, specimen failures and rework.

Three alternatives were tested:

• add a second accessioning clerk during the two-hour peak;
• retain one clerk but stagger transport arrivals;
• add a clerk and redesign priority rules.

The model was verified through event tracing and validated against four weeks of observed arrival and turnaround data. Ten independent replications were run for each alternative after a warm-up period. Results indicated that staggering transport reduced the 90th-percentile turnaround time substantially, while the additional clerk produced a smaller improvement at higher recurring cost. The combined alternative performed best but had diminishing returns because testing capacity then became the bottleneck.

Solution and outcome

Management adopted staggered transport windows, a documented urgent-specimen rule and temporary peak staffing on high-volume days. A dashboard displayed arrival rate, accessioning queue, testing queue, 90th-percentile turnaround time and urgent-specimen compliance. After implementation, the laboratory monitored results for eight weeks and reviewed whether the model assumptions remained valid.

Learning points

The case demonstrates that a local capacity increase may shift rather than remove a bottleneck. Queueing calculations provided an initial diagnosis, but simulation was needed to represent batching, priorities and multiple resources. The recommended action emerged from cost, service performance and operational feasibility—not from a single average waiting-time measure.

10. Best practices and limitations

Best practice begins with a precise decision statement, not with a preferred analytical technique. Analysts should define the population, time horizon, objective, constraints and decision owner. Data should be cleaned without removing meaningful unusual events. Units, denominators and time bases should be explicit.

Models should be proportionate. A spreadsheet may be sufficient for a small production-mix problem; a queueing formula may be useful for rapid screening; simulation may be justified for interacting and stochastic systems. Independent review is valuable for high-consequence decisions.

Common limitations include poor data quality, changing processes, omitted constraints, false assumptions of independence, inappropriate probability estimates and optimisation of a metric that does not represent customer or organisational value. Models can also encourage short-term efficiency at the expense of resilience, learning, safety or employee wellbeing.

A sound decision process therefore combines quantitative evidence with process observation, expert judgement, stakeholder consultation and post-implementation learning. The recommendation should include assumptions, expected benefits, risks, monitoring indicators and trigger points for revising the decision.

Chapter Summary

Operations analytics progresses from descriptive understanding to diagnosis, prediction and prescription. Descriptive measures summarise performance; predictive models estimate future outcomes; prescriptive models compare actions under constraints and uncertainty. Decision trees and expected monetary value support choices under risk, while expected opportunity loss and value of information clarify the worth of additional knowledge.

Linear programming expresses objectives and resource constraints mathematically and can reveal capacity values through shadow prices. Queueing models explain why high utilisation creates disproportionate waiting, although their assumptions must be checked. Simulation represents complex systems with interacting variability and requires verification, validation and replication. Sensitivity and scenario analysis test the stability of recommendations. Dashboards translate analysis into timely operational action, but must use meaningful definitions, targets and escalation rules.

Effective analytics is not merely technical. It requires sound data governance, ethical oversight, transparent assumptions, human accountability and continuous review. The strongest operations decisions combine models with process knowledge and disciplined learning.

Key Terms

Descriptive analytics: Analysis that summarises historical or current performance.

Diagnostic analytics: Analysis used to investigate causes of observed outcomes.

Predictive analytics: Estimation of future outcomes using data and models.

Prescriptive analytics: Evaluation of actions to recommend decisions.

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

Expected monetary value: Probability-weighted average payoff of an alternative.

Expected opportunity loss: Probability-weighted regret from not selecting the best action after the state is known.

Value of information: Expected benefit of obtaining additional information before deciding.

Linear programming: Optimisation with linear objectives and constraints.

Decision variable: A controllable quantity selected by an optimisation model.

Constraint: A limitation or required relationship in a decision model.

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

Queue: Jobs or customers waiting for or receiving service.

Utilisation: Proportion of available service capacity being used.

Little’s Law: Relationship L = λW between average number, arrival rate and time in a stable system.

Simulation: Computer-based imitation of system behaviour over time.

Verification: Checking whether a model has been implemented correctly.

Validation: Checking whether a model adequately represents the real system for its purpose.

Sensitivity analysis: Examination of output changes when assumptions change.

Dashboard: Visual decision-support display of selected operational measures.

Review Questions

Short Answer Questions

1. Distinguish descriptive, predictive and prescriptive analytics.
2. State the stability condition for an M/M/1 queue and explain its meaning.
3. What is the difference between a decision node and a chance node in a decision tree?
4. Define a shadow price in linear programming.
5. Why are verification and validation both required in simulation?

Long Answer Questions

1. Explain how an operations analyst should move from a business problem to a prescriptive model, including decisions, objectives, constraints and uncertainty.
2. Discuss the assumptions, formulas and managerial interpretation of the M/M/1 queueing model.
3. Explain how sensitivity analysis and scenario analysis improve the quality of operational decisions.
4. Compare linear programming and discrete-event simulation, including appropriate applications and limitations.
5. Evaluate the governance and ethical risks of using AI for predictive maintenance or quality inspection.

Analytical / Application Questions

1. A service desk receives 12 requests per hour and can serve 15 per hour. Using the M/M/1 model, calculate utilisation and expected time in the system. Interpret the result and identify assumptions that should be checked.
2. A firm must choose between two capacity plans under high and low demand. Construct a decision tree, calculate expected monetary value for each plan, and explain how the decision could change if probabilities were uncertain.
3. Formulate a linear programming model for a two-product process with labour, machine and demand constraints. Explain which sensitivity results would be most useful to the operations manager.

Multiple Choice Questions with Answers

Question 1

Which type of analytics recommends an action subject to capacity and budget constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

An M/M/1 system has arrival rate λ = 8 per hour and service rate μ = 10 per hour. What is its utilisation?

A. 0.20

B. 0.80

C. 1.25

D. 2.00

Correct Answer: B. 0.80

Question 3

For the production-mix example in the chapter, which solution satisfies both binding capacity constraints?

A. S = 50, P = 0

B. S = 0, P = 40

C. S = 40, P = 20

D. S = 20, P = 40

Correct Answer: C. S = 40, P = 20

Question 4

What does a shadow price normally estimate?

A. The average selling price of a product

B. The marginal value of relaxing a resource constraint within a valid range

C. The probability of a machine failure

D. The total fixed cost of a process

Correct Answer: B. The marginal value of relaxing a resource constraint within a valid range

Question 5

If λ = 8 per hour and μ = 10 per hour in an M/M/1 queue, what is the expected time in the system?

A. 0.05 hour

B. 0.10 hour

C. 0.50 hour

D. 1.25 hours

Correct Answer: C. 0.50 hour

Question 6

Which statement best describes simulation validation?

A. Checking whether the computer code runs without syntax errors

B. Checking whether the model adequately represents the real system for its intended purpose

C. Selecting the alternative with the highest payoff

D. Increasing the number of dashboard indicators

Correct Answer: B. Checking whether the model adequately represents the real system for its intended purpose

Question 7

The expected value of perfect information is calculated as:

A. Best EMV without information minus worst payoff

B. Expected payoff with perfect information minus best EMV without it

C. Total cost divided by total demand

D. Probability multiplied by utilisation

Correct Answer: B. Expected payoff with perfect information minus best EMV without it

Question 8

Which condition is required for the basic M/M/1 queue to be stable?

A. ρ = 0

B. ρ < 1

C. ρ > 1

D. λ = μ + 1

Correct Answer: B. ρ < 1

Question 9

What is the primary purpose of sensitivity analysis?

A. To eliminate all uncertainty from a model

B. To determine how conclusions change when assumptions or inputs change

C. To replace process observation

D. To guarantee an optimal implementation result

Correct Answer: B. To determine how conclusions change when assumptions or inputs change

Question 10

Which dashboard practice is most appropriate?

A. Display every available measure regardless of purpose

B. Use undefined traffic-light thresholds

C. Link a small set of defined measures to decisions and escalation rules

D. Report averages without denominators or time windows

Correct Answer: C. Link a small set of defined measures to decisions and escalation rules