Supply Chain Analytics and Decision Models

Introduction

Global supply chains generate data on orders, inventory, production, transportation, customs clearance, suppliers, customers and disruptions. Analytics converts these observations into decisions. Descriptive analytics explains what happened; predictive analytics estimates what may happen; and prescriptive analytics recommends actions subject to operational constraints.

Supply chain decisions are interconnected. A lower-cost sourcing decision may increase lead-time risk, while inventory reduction may worsen service performance. Decision models therefore need explicit objectives, constraints, uncertainty assumptions and data-quality controls. Model outputs should inform managerial judgment rather than replace it.

Learning Objectives

After studying this chapter, students should be able to:

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Calculate and interpret logistics performance measures and forecast errors.
3. Develop cost-to-serve models for customers, products and channels.
4. Formulate and solve a basic supply network optimization model.
5. Apply simulation, scenario analysis and sensitivity analysis.
6. Validate constraints and assess model feasibility.
7. Evaluate data, ethical and governance limitations in supply chain analytics.

1. The Supply Chain Analytics Framework

Analytics can be organized by the question being answered.

These categories form a decision cycle rather than isolated stages. Forecasts become inputs to optimization; actual outcomes are then measured and used to revise models.

Figure: Analytics forms a continuous measurement, prediction, decision and learning cycle.

2. Descriptive Analytics and Performance Measurement

A performance measure is useful only when its definition, unit, time period and data source are consistent. Common logistics indicators include:

• order cycle time;
• on-time-in-full delivery (OTIF);
• inventory turnover;
• transport cost per shipment or tonne-kilometre;
• warehouse cost per order line;
• forecast accuracy;
• damage and return rates;
• carbon emissions per shipment.

If an order counts as successful only when it is both on time and complete, then:

The numerator counts qualifying orders and the denominator counts orders governed by the agreed service standard. Cancelled orders or customer-requested delays must be treated according to a documented policy.

Segmentation and diagnostic analysis

Aggregate averages can hide important variation. Analysts should segment data by customer, product, country, carrier, route, supplier and service level. A mean transit time of eight days may combine a reliable six-day route with an unstable ten-day route.

Diagnostic analysis investigates causal mechanisms, although observational associations do not automatically establish causation. A correlation between port congestion and late delivery may be operationally informative, but other factors—such as weather or vessel schedules—may influence both variables.

3. Cost-to-Serve Analysis

Cost-to-serve (CTS) estimates the resources consumed in serving a customer, product, order, channel or market. Traditional accounting averages may conceal the high cost of fragmented orders, special packaging, remote delivery or frequent returns.

Under activity-based costing, total service cost for object k is:

where r_a is the cost-driver rate for activity a in currency per driver unit, and q_ak is the quantity of that driver consumed by cost object k. The model assumes activities and drivers have been defined consistently and that allocated capacity costs are relevant to the decision.

Worked example: customer-order profitability

An international order generates revenue of $100 and has product cost of $40. Its logistics activities are:

The order contribution is:

Thus, contribution equals $100 − $40 − $44 = $16, or 16% of revenue. CTS supports pricing, service differentiation and process improvement. It should not be used mechanically to abandon customers whose current costs may be temporary or strategically justified.

4. Predictive Analytics for Logistics Decisions

Predictive methods estimate uncertain quantities such as demand, lead time, transport delay, supplier failure or return probability. Methods range from moving averages and exponential smoothing to regression, classification and machine learning.

Forecast error measures

For period t, define forecast error as:

where A_t is actual demand and F_t is forecast demand, both in the same units.

Common measures are:

MAE is measured in the original demand units. RMSE penalizes large errors more heavily. MAPE is undefined when an actual value is zero and can become unstable when actual values are very small.

For actual demands of 100, 120 and 80 units and forecasts of 90, 125 and 100 units, errors are 10, −5 and −20 units. Therefore:

• MAE = (10 + 5 + 20) ÷ 3 = 11.67 units;
• RMSE = √[(100 + 25 + 400) ÷ 3] = √175 ≈ 13.23 units;
• MAPE = (10% + 4.17% + 25%) ÷ 3 ≈ 13.06%.

Forecast evaluation must use data not employed in model fitting. Using future information in predictors creates data leakage and gives misleadingly optimistic performance.

Predictive models and decision relevance

The most statistically accurate model is not necessarily the most valuable. Underestimating demand for critical medical products may be more costly than overestimating it. Model selection should therefore consider asymmetric business consequences, interpretability and robustness under structural change.

5. Prescriptive Analytics and Optimization

Optimization selects decision-variable values that maximize or minimize an objective while satisfying constraints. Typical objectives include cost, profit, lead time, emissions or a weighted combination of criteria.

General distribution-network formulation

Let:

• i ∈ I index plants or warehouses;
• j ∈ J index markets;
• x_ij be units shipped from facility i to market j;
• c_ij be variable shipping cost in currency per unit;
• K_i be facility capacity in units per planning period;
• D_j be market demand in units per planning period.

For fixed, operating facilities, the transportation model is:

subject to:

The first constraints enforce capacities; the second satisfy each market exactly. If unmet demand is permitted, shortage variables and penalty costs must be explicitly introduced rather than silently violating demand constraints.

For facility-location decisions, define binary y_i = 1 when facility i is open and 0 otherwise, and fixed cost f_i. The objective adds ∑_i f_i y_i, while capacity becomes ∑_j x_ij ≤ K_i y_i. This is a mixed-integer model.

The basic formulation assumes divisible products, known costs and demands, compatible capacities, and no unmodelled route restrictions. Product, time, emissions, tariffs or service constraints can be added through further indices and equations.

Complete worked network model

Two plants serve two markets. Costs are dollars per unit.

Market demands are 80 units for M1 and 100 units for M2. The model is:

subject to:

and all variables non-negative.

The low-cost allocation is x₁₁=80, x₁₂=0, x₂₁=0 and x₂₂=100. Its cost is:

Every constraint is satisfied: P1 uses 80 of 100 units, P2 uses 100 of 120, M1 receives 80 and M2 receives 100.

A feasible alternative is x₁₁=60, x₁₂=40, x₂₁=20 and x₂₂=60. It meets both demands and uses 100 units at P1 and 80 at P2. Its cost is $760, so it is feasible but not optimal.

By contrast, the candidate allocation x₁₁=70, x₁₂=40, x₂₁=10, x₂₂=70 is infeasible for two separate reasons. P1 ships 110 units, exceeding its 100-unit capacity, and M2 receives 110 units rather than its required 100. M1 and P2 happen to satisfy their constraints, but partial compliance does not make a solution feasible. Analysts must check every constraint.

Figure: Optimal flows use P1 for M1 and P2 for M2; dashed routes remain available but unused.

6. Sensitivity Analysis

Sensitivity analysis examines how results change when an input changes. It helps managers understand whether a recommendation is robust or dependent on uncertain assumptions.

In the two-by-two network, shifting one unit from the direct-route pattern P1–M1 and P2–M2 to the cross-route pattern P1–M2 and P2–M1 changes cost by:

For the example, ∆Z=(6+5)−(4+3)=$4 per exchanged unit. Therefore, the direct-route pattern is cheaper while c₁₁+c₂₂<c₁₂+c₂₁, provided capacities and demand permit the exchange. Equality would create alternative optimal allocations along that cycle.

A shadow price, when available from a linear-programming solution, measures the marginal objective improvement from relaxing a binding constraint by one unit within an allowable range. It is not necessarily valid for large capacity changes or integer models.

Good sensitivity analysis varies influential inputs—including demand, fuel cost, exchange rates, tariffs and capacity—individually and jointly. One-way analysis is easy to interpret but may overlook interactions.

7. Scenario Analysis

Scenario analysis evaluates coherent combinations of assumptions. Unlike a forecast, a scenario is not necessarily the most likely future. Unlike Monte Carlo simulation, it normally examines a limited set of deliberately constructed conditions.

For the network example:

If an emergency supplier has unlimited capacity and costs $9 to M1 and $8 to M2, the disruption solution can send 80 units from P1 to M1, 20 from P1 to M2, 70 from P2 to M2 and 10 from the emergency supplier to M2. The total is:

The scenario exposes the value of backup capacity. It also shows why merely changing demand without re-optimizing can produce infeasible plans.

8. Simulation Under Uncertainty

Optimization commonly uses point estimates. Simulation represents uncertain inputs with probability distributions and repeatedly evaluates outcomes. Monte Carlo simulation is useful when analytical solutions are difficult or when managers need distributions of cost, delay or service—not only expected values.

A defensible simulation follows five steps:

1. Define the decision alternatives and performance measures.
2. Specify input distributions from evidence and operational knowledge.
3. Generate random input realizations.
4. calculate outcomes for each replication.
5. Summarize means, percentiles, variability and confidence intervals.

Worked Monte Carlo inventory simulation

A distributor must choose an order quantity of either 100 or 120 units for a selling period. Unsold inventory costs $2 per unit and shortage costs $8 per unit. Demand has this discrete distribution:

For order quantity Q and simulated demand D, period cost is:

This equation assumes no revenue difference between alternatives, no replenishment during the period and linear overage and shortage costs.

Using random numbers 07, 42, 68, 91, 24, 76, 55, 13, 88 and 36 produces demands 70, 90, 110, 130, 90, 110, 110, 70, 130 and 90.

The simulation favors ordering 120 units, with an estimated average cost of $60 rather than $90. Common random numbers were used for both alternatives, enabling comparison under identical demand realizations. Ten trials are insufficient for most professional decisions; many replications and convergence checks are normally required. In this illustration, the generated frequencies exactly match the stated probabilities, so the simulated means also equal the exact expected costs.

Simulation does not optimize automatically. It evaluates alternatives supplied by the analyst unless combined with a search or optimization procedure. Poor input distributions will still produce misleading results.

9. Illustrative Applied Case: Aster Home Products

Background

Aster Home Products is a fictional European consumer-goods company supplying two overseas markets from two plants. Rapid international growth has increased freight expenditure, while planners rely on spreadsheets and average customer margins.

Problem

Management needs to determine plant-to-market allocations, assess disruption exposure and identify customers whose service requirements erode profitability. Data come from enterprise-resource-planning, warehouse, carrier and customs systems, but customer identifiers and currency conversions are inconsistent.

Analysis and approach

The project team first establishes a common data dictionary and converts all costs to US dollars using documented period exchange rates. It then:

1. builds activity-based CTS profiles using picking, packaging, transport, customs, returns and support drivers;
2. validates demand and capacity data with operations managers;
3. formulates the complete network model developed earlier;
4. obtains the $620 base allocation and verifies every capacity and demand constraint;
5. tests demand-surge and P2-disruption scenarios;
6. simulates inventory alternatives under uncertain demand;
7. compares recommendations with service, contractual and geopolitical considerations.

Outcome

The model recommends the direct-route allocation in the base case and identifies emergency supply as necessary when P2 capacity falls to 70 units. CTS analysis shows that small international orders are costly primarily because of handling and customs activities, not product cost. Management therefore considers consolidated ordering and revised service options rather than indiscriminate customer withdrawal.

Learning points

• Optimization must be preceded by data reconciliation and followed by constraint validation.
• A low-cost plan may be fragile under disruption.
• Scenario analysis, simulation and CTS answer different but complementary questions.
• Managerial judgment remains necessary where contracts, relationships or social consequences are not fully represented in the objective function.

10. Data Limitations and Model Risk

Supply chain data are often incomplete, delayed or semantically inconsistent. A field labelled “delivery date” might mean requested, promised or actual delivery date. Common limitations include:

• missing or duplicated transactions;
• inconsistent product and location codes;
• inaccurate lead-time timestamps;
• selection bias from recording only completed shipments;
• censored demand when stockouts hide unmet demand;
• changing processes and structural breaks;
• currency, tax and unit-conversion errors;
• weak master-data governance.

Model risk also arises from incorrect assumptions. Linear costs may ignore volume discounts; normal demand distributions may assign unrealistic probabilities; and historical relationships may fail during conflict, pandemics or regulatory change. Models should be validated against historical cases, tested under extreme conditions and monitored after deployment.

11. Governance, Ethics and Responsible AI

Analytics can affect suppliers, employees, carriers and customers. Cost-minimizing systems may shift unrealistic workloads to drivers, discriminate indirectly through location variables or favor suppliers with incomplete environmental reporting.

Responsible practice requires:

• clearly assigned ownership for data and models;
• role-based access and cybersecurity controls;
• compliance with privacy, competition, customs and trade regulations;
• documentation of assumptions, versions and overrides;
• bias and performance monitoring across affected groups;
• human review of high-impact decisions;
• auditable explanations for recommendations.

Machine learning can improve demand, delay and risk prediction, while digital twins can combine network models with near-real-time data. Generative AI may assist query construction or exception summaries, but it should not be trusted to invent operational facts, approve shipments or modify optimization constraints without verification.

12. Managerial Best Practices

Effective supply chain analytics projects generally:

1. begin with a decision rather than an available dataset;
2. define objectives and constraints with decision owners;
3. establish units, time horizons and data definitions;
4. maintain separate training, validation and test data for prediction;
5. compare models against simple benchmarks;
6. test feasibility before comparing objective values;
7. use sensitivity, scenarios and simulation to expose uncertainty;
8. document assumptions and create override procedures;
9. monitor realized business and service outcomes;
10. revise models when markets, regulations or processes change.

Chapter Summary

Supply chain analytics links operational data to managerial decisions. Descriptive and diagnostic methods measure performance and investigate variation; predictive models estimate uncertain demand, lead time and risk; prescriptive models recommend actions subject to constraints. Cost-to-serve analysis traces activity costs to customers and orders, while network optimization determines feasible allocations across facilities and markets.

A valid optimization solution must satisfy every capacity, demand, flow and non-negativity constraint. Sensitivity analysis tests input changes, scenario analysis compares coherent futures, and simulation evaluates repeated uncertain outcomes. These methods are complementary rather than interchangeable. Their usefulness depends on appropriate assumptions, reliable data, careful validation, ethical governance and informed human judgment.

Key Terms

Activity-based costing: Assignment of resource costs using measurable activity drivers.

Constraint: A mathematical restriction that a decision solution must satisfy.

Cost-to-serve: Total activity cost incurred in serving a customer, product, order or channel.

Data leakage: Use of information in model development that would not be available at prediction time.

Descriptive analytics: Analysis explaining what has occurred.

Diagnostic analytics: Analysis investigating why an outcome occurred.

Digital twin: A dynamic digital representation of a physical process or network.

Feasible solution: A set of decision-variable values satisfying every model constraint.

Forecast error: Difference between an actual value and its forecast.

Monte Carlo simulation: Repeated sampling from probability distributions to estimate outcome distributions.

Network optimization: Mathematical selection of flows, capacities or locations across connected nodes and routes.

Objective function: Mathematical expression to be minimized or maximized.

Prescriptive analytics: Analysis recommending actions under stated objectives and constraints.

Predictive analytics: Analysis estimating future or unknown outcomes.

Scenario analysis: Evaluation of decisions under coherent sets of alternative assumptions.

Sensitivity analysis: Assessment of how results change when model inputs vary.

Shadow price: Marginal objective change associated with relaxing a binding linear constraint within a valid range.

Simulation: Experimental representation of a system’s behavior over repeated or evolving conditions.

Review Questions

Short Answer Questions

1. Distinguish descriptive analytics from prescriptive analytics.
2. Why can MAPE be unsuitable when actual demand is zero or very small?
3. What does it mean for a network solution to be feasible?
4. How does scenario analysis differ from Monte Carlo simulation?
5. Define cost-to-serve and identify two possible cost drivers.

Long Answer Questions

1. Explain how descriptive, predictive and prescriptive analytics interact in supply chain planning.
2. Discuss the assumptions and limitations of a transportation optimization model.
3. Explain how simulation should be designed, validated and interpreted in an inventory decision.
4. Evaluate the managerial benefits and risks of customer-level cost-to-serve analysis.
5. Discuss data governance and ethical requirements for AI-enabled supply chain analytics.

Analytical / Application Questions

1. A facility has capacity of 150 units and proposed shipments of 60, 55 and 45 units. Determine feasibility and explain the required correction.
2. Develop a two-facility, three-market transportation model, defining all variables, objective terms and constraints.
3. Design three scenarios for an international distribution network exposed to fuel-price, exchange-rate and port-disruption risks, and explain which outputs management should compare.

Multiple Choice Questions with Answers

Question 1

Which type of analytics recommends a shipment allocation subject to capacity and demand constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

Actual demand is 100, 120 and 80 units, while forecasts are 90, 125 and 100. What is the MAE?

A. 10.00 units

B. 11.67 units

C. 13.23 units

D. 35.00 units

Correct Answer: B. 11.67 units

Question 3

In the worked network, consider x₁₁=70, x₁₂=40, x₂₁=10 and x₂₂=70. Which constraints are violated?

A. Only the P1 capacity constraint

B. Only the M1 demand constraint

C. The P1 capacity and M2 demand constraints

D. The P2 capacity and M1 demand constraints

Correct Answer: C. The P1 capacity and M2 demand constraints

Question 4

What is the minimum transportation cost in the worked two-plant, two-market network?

A. $620

B. $680

C. $720

D. $760

Correct Answer: A. $620

Question 5

An order has picking, packing, transport, customs, expected-return and support costs of $12, $4, $18, $3, $1 and $6. What is its cost-to-serve?

A. $38

B. $40

C. $42

D. $44

Correct Answer: D. $44

Question 6

In the ten-trial simulation, what was the average cost when Q=120 units?

A. $40

B. $60

C. $80

D. $90

Correct Answer: B. $60

Question 7

Which statement correctly distinguishes scenario analysis from Monte Carlo simulation?

A. Scenario analysis normally tests selected coherent futures, while Monte Carlo simulation repeatedly samples uncertain inputs.

B. Scenario analysis always assigns probabilities, while simulation never uses probabilities.

C. Scenario analysis optimizes automatically, while simulation guarantees feasibility.

D. Scenario analysis requires historical data, while simulation cannot use historical data.

Correct Answer: A. Scenario analysis normally tests selected coherent futures, while Monte Carlo simulation repeatedly samples uncertain inputs.

Question 8

For the two-by-two network, when is the direct-route pattern cheaper than exchanging flow to the two cross-routes, assuming the exchange is feasible?

A. c₁₁+c₂₂<c₁₂+c₂₁

B. c₁₁+c₂₂>c₁₂+c₂₁

C. c₁₁=c₁₂ regardless of other costs

D. c₂₁=c₂₂ regardless of other costs

Correct Answer: A. c₁₁+c₂₂<c₁₂+c₂₁

Question 9

Which practice creates data leakage in a demand-forecasting model?

A. Evaluating forecasts with MAE

B. Using future sales information that would not exist at prediction time

C. Comparing a complex model with a simple benchmark

D. Separating training and test datasets chronologically

Correct Answer: B. Using future sales information that would not exist at prediction time

Question 10

Which control best supports responsible use of prescriptive supply chain models?

A. Removing all human review

B. Keeping assumptions and overrides undocumented

C. Documenting model versions, constraints and authorized overrides

D. Optimizing cost without monitoring service outcomes

Correct Answer: C. Documenting model versions, constraints and authorized overrides