Freight Analytics and Decision Models

Introduction and Context

International freight decisions are no longer made solely through tariff tables, carrier quotations, or the experience of a shipping manager. Contemporary logistics managers must combine shipment data, market forecasts, network constraints, service commitments, emissions information, geopolitical risk, and financial objectives. Freight analytics provides the methods for converting these data into decisions about routing, transport mode, capacity, demand, inventory, and total landed cost.

Three analytical questions are particularly important:

1. What has happened? Descriptive analytics measures and explains historical performance.
2. What is likely to happen? Predictive analytics estimates future demand, transit time, disruption probability, and cost.
3. What should be done? Prescriptive analytics recommends decisions subject to commercial, operational, legal, and resource constraints.

A sound freight decision model does not treat the cheapest quoted freight rate as the lowest-cost option. It evaluates the total consequences of a choice, including inventory in transit, handling, duties, delay exposure, carbon costs, reliability, and the value of flexibility.

Learning Objectives

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

• distinguish descriptive, predictive, and prescriptive freight analytics;
• construct a total-cost model for international transport;
• compare modes and routes using multicriteria analysis;
• apply forecasting, simulation, and optimization appropriately;
• interpret capacity, demand, cost, and service trade-offs;
• perform sensitivity and scenario analysis;
• design managerial dashboards and governance controls; and
• evaluate the ethical, regulatory, and practical limitations of freight analytics.

1. The Analytical Architecture of Freight Decisions

A freight analytics system normally contains five connected layers.

The layers should not be confused. A dashboard showing that air freight was expensive is descriptive; a model predicting that a port strike will increase transit time is predictive; a recommendation to divert selected shipments by rail is prescriptive.

Data foundations

Important freight data include purchase orders, shipment dates, origin and destination, product dimensions and weight, Incoterms, carrier and service level, quoted and invoiced charges, customs records, inventory levels, booking cancellations, estimated and actual arrival times, and disruption events. Data should be reconciled across systems because a transport-management system may record a booking date while an enterprise-resource-planning system records the invoice date.

A useful data-quality framework examines completeness, accuracy, timeliness, consistency, uniqueness, and lineage. For example, a predictive model trained on invoices that omit accessorial charges will systematically understate future transport cost. Data leakage must also be prevented: information known only after delivery must not be used to predict the original delivery outcome.

2. Descriptive Analytics: Establishing the Baseline

Descriptive analytics converts historical transactions into measures that reveal performance and variation. Common measures include freight spend, cost per shipment, cost per kilogram, cost as a percentage of sales, lead time, transit-time variability, tender acceptance, damage, claims, customs dwell, and on-time-in-full delivery.

For a set of shipments, the mean cost is:

where C_i is the cost of shipment i, n is the number of shipments, and C̄ is the arithmetic mean in the currency used. The mean is sensitive to exceptional shipments, so the median and percentiles should also be reported.

A weighted average cost per kilogram is more informative than the simple average of shipment-level ratios:

where W_i is the chargeable weight of shipment i. Units are currency per kilogram. This measure is appropriate when shipment weights differ materially.

For delivery reliability, an on-time rate may be calculated as:

The definition of “on time” must be fixed in advance, including time zone, calendar days, partial deliveries, and customer-specific tolerance.

Descriptive analysis should examine distributions, not merely averages. A lane with a mean transit time of ten days may be preferable to a lane with a mean of eight days if the first has much lower variance and fewer extreme delays. Pareto analysis can identify the small number of lanes, suppliers, products, or accessorial charges responsible for a large share of cost or service failures.

3. Predictive Analytics for Demand, Transit and Risk

Demand forecasting

Demand forecasts support booking, capacity reservations, inventory positioning, and workforce planning. A simple moving average is:

where D_t is observed demand in period t, k is the number of periods included, and F_t+1 is the forecast for the next period. This method assumes that recent observations are representative and gives equal weight to each included period.

Exponential smoothing gives greater weight to recent observations:

where 0 ≤ α ≤ 1. A high α reacts quickly to changes but may overreact to noise; a low α is more stable but slower to recognize a structural shift. Seasonal demand requires seasonal methods or explanatory variables rather than an unadjusted moving average.

Forecast accuracy should be assessed out of sample. Mean absolute error is:

and root mean squared error is:

MAE has the same units as demand and is easy to interpret. RMSE penalizes large errors more heavily. No metric is universally superior; the selected metric should reflect the cost of forecast errors.

Transit-time and disruption prediction

Transit prediction may use lane, carrier, port, season, shipment type, weather, congestion, customs complexity, and historical schedule adherence. Classification models can estimate the probability of a late arrival, while regression models estimate transit duration. Predictive accuracy must be separated from decision usefulness: a slightly less accurate model may be preferable if it is more explainable, stable, and operationally actionable.

Probability estimates should be calibrated. If shipments assigned a 20% delay probability experience delays approximately 20% of the time, the model is calibrated. Managers should also distinguish aleatory uncertainty, which is inherent randomness, from epistemic uncertainty, which arises from limited knowledge or poor data.

4. Total-Cost Models

A transport mode should be evaluated using total relevant cost rather than the freight quote alone. A general shipment-level model is:

where:

• TC = total relevant cost;
• F = line-haul freight and fuel-related charges;
• H_o and H_d = origin and destination handling costs;
• C_c = customs, documentation, and compliance cost;
• D = duties and import taxes, where relevant;
• I_t = inventory carrying cost during transit;
• R = expected cost of delay, damage, or disruption;
• E = relevant environmental or emissions cost; and
• Q = other quality, security, or special-service costs.

The equation is a decision model, not an accounting identity. Terms should be included only when they are incremental, measurable, and relevant to the alternatives being compared. Sunk costs should not influence a forward-looking choice.

If average inventory value in transit is V, annual carrying-cost rate is r, and transit time is t days, then a simple in-transit inventory cost is:

This assumes a constant annual carrying rate, a 365-day year, and that the relevant inventory value remains exposed throughout the transit period. If inventory value changes, or if financing and obsolescence have different time profiles, the model should be expanded.

Expected delay cost can be represented as:

where p_d is the probability of a defined delay event and L_d is its monetary loss. The loss may include stockout margin, production interruption, contractual penalties, expediting, and customer recovery. Expected value is useful for repeated decisions but does not describe the full distribution of severe outcomes; risk-averse managers may therefore add a service constraint or tail-risk measure.

Worked example: mode comparison

A firm compares ocean and air transport for goods with an in-transit value of $120,000. The annual carrying rate is 18%. Ocean transit is 28 days and air transit is 5 days. Ocean freight and associated charges are $8,400; air freight and associated charges are $29,000. The estimated expected disruption costs are $3,600 for ocean and $900 for air. Ignore duties because they are identical.

Ocean inventory cost is:

I_ocean = 120,000 × 0.18 × (28 ÷ 365) = $1,656.99

Thus, ocean total relevant cost is approximately $13,656.99. Air inventory cost is:

I_air = 120,000 × 0.18 × (5 ÷ 365) = $295.89

Air total relevant cost is approximately $30,195.89. Ocean is cheaper by about $16,538.90 under these assumptions. However, the conclusion changes if the shipment is urgently needed or if the expected disruption loss has been understated. The calculation therefore supports, rather than replaces, managerial judgment.

5. Multicriteria Decision Analysis

Cost minimization is inappropriate when service, resilience, safety, emissions, and strategic relationships matter. Multicriteria decision analysis (MCDA) makes these trade-offs explicit.

In a weighted-sum model, alternative j receives score:

where w_m is the weight of criterion m, z_jm is the normalized score of alternative j on criterion m, and the weights sum to one. Scores must be oriented consistently: a higher normalized score should always mean more desirable performance.

For a beneficial criterion, min–max normalization can be:

For a cost or risk criterion, reverse the direction:

Normalization is sensitive to extreme alternatives. If the decision is strategically important, managers should test alternative normalization methods and weights. MCDA should not conceal value judgments: a weight of 40% for cost and 30% for reliability is a managerial policy choice, not a discovered fact.

A practical model may include cost, transit time, reliability, emissions, customs capability, network flexibility, and geopolitical exposure. Some criteria should be treated as constraints rather than tradeable scores. For example, a carrier lacking a required dangerous-goods certification should be excluded, regardless of its low price.

6. Simulation and Scenario Analysis

Simulation is valuable when freight outcomes are uncertain and analytical formulas are insufficient. A Monte Carlo model repeatedly samples uncertain inputs—such as demand, transit time, fuel price, port dwell, or disruption occurrence—and calculates the resulting cost or service outcome.

A generic simulation procedure is:

1. define the decision and performance measures;
2. identify uncertain inputs and estimate their distributions from evidence;
3. generate a random or quasi-random scenario;
4. calculate cost, service, capacity use, and emissions;
5. repeat the experiment many times;
6. summarize means, percentiles, probabilities, and downside outcomes; and
7. compare alternatives using the same scenarios where possible.

Correlations must be represented. Port congestion and transit delay, for example, are not independent. Ignoring correlation can produce an unrealistically optimistic risk estimate. Simulation outputs should include a confidence interval for estimated probabilities where appropriate, and input assumptions should be documented.

Scenario analysis differs from simulation. A scenario is a coherent narrative or parameter set—such as “normal operations,” “fuel-price shock,” or “major port closure.” It is particularly useful for strategic stress testing, where historical frequencies may not represent future structural change.

7. Optimization Models for Routing and Capacity

Optimization selects values for decision variables to minimize or maximize an objective subject to constraints. A simplified shipment-routing model can be written as:

subject to:

Here, x_ijk is the quantity shipped from origin i to destination j using mode or route k; c_ijk is unit cost; d_i is available demand or supply; and s_j is capacity. Units must be consistent, such as tonnes and currency per tonne. Additional constraints can represent minimum order quantities, carrier allocations, customs restrictions, lane availability, or service requirements.

If a route is either selected or not selected, a binary variable is required. Mixed-integer programming is then appropriate, although computational difficulty may increase. A model is only as good as its constraints: omitting a booking deadline or port capacity may produce an “optimal” solution that cannot be executed.

Objective design

A single objective may minimize total cost, but a more realistic objective can combine cost and penalties for lateness, emissions, and unserved demand. Alternatively, managers can use a constrained objective: minimize cost subject to on-time delivery of at least 95%, emissions below a target, and no more than a specified share assigned to one carrier. Constrained formulations often communicate policy more clearly than arbitrary monetary penalties.

8. Sensitivity, Robustness and Value of Information

Sensitivity analysis asks how the recommendation changes when assumptions change. Key variables include fuel price, exchange rates, demand volume, delay probability, carrying-cost rate, accessorial charges, and capacity availability.

A one-way sensitivity analysis changes one variable at a time. A tornado chart ranks variables by their effect on the outcome. Multiway analysis changes related variables together, while stress testing examines severe but plausible conditions. Break-even analysis identifies the value at which two alternatives become equally attractive.

For the mode example, the air–ocean break-even expected disruption cost for ocean can be found by equating totals. If all other ocean and air values remain as above, ocean becomes equal to air when ocean’s disruption cost is approximately $20,138.90 rather than $3,600. This illustrates the importance of validating risk estimates before changing mode.

Robust decisions perform acceptably across a range of futures rather than optimizing one forecast. Robust optimization may protect against uncertainty sets, while managerial practice may use capacity buffers, dual sourcing, or contingent routing. The best choice is not necessarily the one with the lowest expected cost; it may be the one with acceptable downside and rapid recovery.

Value of information measures whether better information is worth acquiring. For example, a paid port-congestion forecast is justified if its expected reduction in disruption and expediting costs exceeds its price, after considering implementation and decision costs.

9. Dashboards and Managerial Communication

A dashboard should support a decision, not merely display data. A useful freight dashboard may contain:

• spend versus budget and forecast;
• cost per kilogram and cost per shipment;
• on-time delivery and transit-time percentiles;
• forecast demand and forecast error;
• booked, available, and utilized capacity;
• exceptions by lane, supplier, carrier, and customer;
• carbon emissions per tonne-kilometre where measurement is credible; and
• current disruption exposure and recommended actions.

Each metric needs a definition, data owner, refresh frequency, target, and escalation rule. A red indicator without a prescribed action creates attention but not control. Dashboards should allow drill-down from global performance to shipment-level evidence and should distinguish actuals, forecasts, and scenario outputs visually and conceptually.

10. Illustrative Case: Meridian Components’ Asia–Europe Lane

Background. Meridian Components, a fictional manufacturer, imports electronic assemblies from three Asian suppliers into a European plant. Demand is seasonal, product obsolescence is significant, and production stoppages are costly. The firm historically used ocean freight because its quoted rate was low, but late deliveries increased emergency air shipments.

Problem. Management must choose a procurement and transport policy for the next quarter. Alternatives are: (A) all ocean; (B) all air; and (C) a hybrid policy using ocean for baseline demand and air for urgent or high-risk orders.

Analysis. The analytics team first reconciles purchase-order, booking, invoice, and production data. It finds that ocean’s average cost is low but its 90th-percentile transit time is substantially longer than its mean. A demand forecast is produced by supplier and week, then tested against a holdout period. Transit-time distributions are estimated by lane and season. The total-cost model includes freight, handling, inventory, expected production interruption, emergency expediting, and emissions. A simulation represents correlated port congestion and seasonal demand.

Approach and solution. The hybrid policy is formulated as a constrained optimization: minimize expected total cost subject to a 97% production-coverage target, a maximum air-freight share, and supplier-specific capacity limits. A scenario test includes a port disruption and a 25% increase in fuel prices. The model recommends reserving ocean capacity for stable baseline demand, pre-authorizing air conversion for orders that cross a risk threshold, and retaining two qualified carriers on the critical lane.

Outcome. In the illustrative analysis, the hybrid policy has a slightly higher expected freight charge than all-ocean shipping but materially lowers expected interruption and expediting cost. It also performs better under the disruption scenario. Management adopts the policy with a weekly exception review and a quarterly refresh of probabilities and cost assumptions.

Learning points. First, the relevant decision is a policy, not a single shipment quote. Second, uncertainty and tail outcomes matter. Third, optimization must include real operational constraints. Fourth, the model’s recommendation is credible only because data definitions, assumptions, and governance are visible.

11. Governance, Ethics and Regulation

Freight analytics can affect carriers, brokers, warehouse workers, suppliers, and customers. Governance therefore includes more than cybersecurity. It requires access control, data minimization, retention rules, model documentation, validation, version control, and an audit trail of decisions.

Algorithmic bias may arise when historical data reflect unequal service allocation, supplier discrimination, or systematic under-recording of delays for powerful customers. A carrier-ranking model should be tested for whether it unfairly penalizes small or new carriers because they have less historical data. Human review is especially important for safety, dangerous goods, sanctions screening, customs classification, and decisions that could terminate a supplier relationship.

Regulatory requirements vary by jurisdiction and product. Managers must consider customs and trade-compliance obligations, sanctions and export controls, dangerous-goods rules, privacy law, competition law, recordkeeping, and environmental reporting standards. Emissions calculations should identify the boundary, activity data, emission factors, and treatment of empty repositioning. False precision in carbon dashboards can create misleading claims.

Generative AI can assist with document classification, exception summaries, and natural-language access to approved data. It should not autonomously approve customs declarations, override safety constraints, or invent missing shipment facts. Retrieval controls, source citations, validation, and human authorization are essential.

12. Limitations and Best Practices

Freight models face structural limitations. Historical relationships may fail during unprecedented disruptions. Costs may be difficult to allocate across shipments. Forecasts can be distorted by promotions, cancellations, and order batching. Optimization may produce mathematically precise but operationally infeasible plans. MCDA weights may reflect politics rather than strategy. Simulation results can be unreliable when distributions are guessed rather than measured.

Best practice is therefore to:

1. define the decision, horizon, scope, and objective before selecting a method;
2. establish a governed data dictionary and reconcile financial and operational records;
3. separate actual, forecast, and scenario values;
4. model constraints and service requirements explicitly;
5. validate forecasts and models out of sample;
6. test sensitivity, stress, correlation, and break-even conditions;
7. compare expected performance with downside exposure;
8. document assumptions, ownership, and model version;
9. monitor drift and recalibrate after structural changes; and
10. combine analytical recommendations with accountable managerial judgment.

Chapter Summary

Freight analytics links data to international logistics decisions through descriptive, predictive, and prescriptive methods. Descriptive measures establish cost, service, and variability baselines. Forecasting estimates demand and transit outcomes, while simulation and scenario analysis represent uncertainty and disruption. Total-cost models broaden evaluation beyond quoted freight to include handling, inventory, delay, quality, emissions, and compliance consequences. MCDA makes trade-offs explicit, and optimization allocates flows subject to demand, capacity, service, and regulatory constraints. Sensitivity and robustness testing reveal whether recommendations survive changed assumptions. Dashboards translate analysis into action when metrics, ownership, definitions, and escalation rules are clear. Finally, data quality, model governance, ethics, and regulatory compliance determine whether a technically sophisticated model can be trusted in practice.

Key Terms

Descriptive analytics: Analysis that summarizes historical or current performance.

Predictive analytics: Analysis that estimates future outcomes or probabilities.

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

Total relevant cost: All incremental cost consequences relevant to comparing alternatives.

Chargeable weight: The weight used to price a shipment under the applicable tariff rule.

On-time delivery: Delivery meeting a defined promised-date standard.

Forecast error: The difference between observed demand and forecast demand.

Mean absolute error: Average absolute forecast error.

Simulation: Repeated evaluation of a model using sampled uncertain inputs.

Scenario analysis: Evaluation of coherent alternative future conditions.

Optimization: Selection of decision variables to improve an objective subject to constraints.

Decision variable: A quantity or choice controlled by the decision-maker.

Constraint: A rule limiting feasible decisions.

Multicriteria decision analysis: Structured comparison using several criteria and weights.

Normalization: Transformation of measures to a common scoring scale.

Sensitivity analysis: Testing how results change when assumptions vary.

Robustness: The ability of a decision to perform acceptably across uncertain conditions.

Model governance: Controls for model ownership, validation, use, monitoring, and auditability.

Data leakage: Use of information unavailable at the time a prediction would have been made.

Review Questions

Short Answer Questions

1. Distinguish descriptive, predictive, and prescriptive freight analytics.
2. Why can weighted cost per kilogram be preferable to a simple average of shipment-level ratios?
3. Identify four components that may be included in a total-cost model.
4. What is the purpose of sensitivity analysis?
5. Why must a freight dashboard define ownership and escalation rules?

Long Answer Questions

1. Explain how an international logistics manager could combine forecasting, simulation, and optimization in a quarterly capacity-planning process.
2. Critically evaluate the assumptions and limitations of a total-cost comparison between air and ocean freight.
3. Describe how MCDA can incorporate service, resilience, cost, and emissions without concealing managerial value judgments.
4. Explain the data, validation, and governance requirements for a predictive late-delivery model.
5. Discuss why a mathematically optimal routing plan may be operationally infeasible.

Analytical / Application Questions

1. A shipment has an in-transit value of $80,000, an annual carrying rate of 15%, and a 24-day transit time. Calculate its in-transit inventory cost using a 365-day year. State the assumptions.
2. Construct a weighted decision matrix for three freight modes using cost, reliability, and emissions. Explain your normalization method and test whether the preferred mode changes when the reliability weight increases.
3. Design a scenario and simulation framework for a lane exposed to seasonal port congestion. Specify uncertain inputs, correlations, performance measures, and the managerial decision rule.

Multiple Choice Questions with Answers

Question 1

Which type of analytics recommends a freight allocation subject to capacity and service constraints?

A. Descriptive analytics

B. Predictive analytics

C. Prescriptive analytics

D. Diagnostic accounting

Correct Answer: C. Prescriptive analytics

Question 2

Which expression correctly calculates weighted cost per kilogram?

A. ∑(Cᵢ ÷ Wᵢ) ÷ n

B. ∑Cᵢ ÷ ∑Wᵢ

C. ∑Wᵢ ÷ ∑Cᵢ

D. (∑Cᵢ × ∑Wᵢ) ÷ n

Correct Answer: B. ∑Cᵢ ÷ ∑Wᵢ

Question 3

A shipment worth $100,000 is in transit for 20 days. At an annual carrying rate of 18%, what is the approximate in-transit inventory cost using 365 days?

A. $493

B. $986

C. $3,600

D. $18,000

Correct Answer: B. $986

Question 4

In a weighted-sum MCDA model, what must the criterion weights generally do?

A. All be greater than one

B. Sum to one when expressed as proportions

C. Be identical for every decision

D. Represent only monetary costs

Correct Answer: B. Sum to one when expressed as proportions

Question 5

What is the principal purpose of Monte Carlo simulation in freight analysis?

A. To eliminate all uncertainty from the data

B. To repeatedly evaluate outcomes under sampled uncertain inputs

C. To replace operational constraints with averages

D. To calculate only historical freight spend

Correct Answer: B. To repeatedly evaluate outcomes under sampled uncertain inputs

Question 6

Which of the following is a constraint in an optimization model?

A. A carrier’s available capacity

B. A randomly generated scenario

C. A dashboard colour

D. A historical median

Correct Answer: A. A carrier’s available capacity

Question 7

Why should port congestion and transit delay often be modelled as correlated?

A. They are both measured in currency

B. Congestion can directly increase delay likelihood

C. Correlation makes every forecast accurate

D. Independent variables cannot be simulated

Correct Answer: B. Congestion can directly increase delay likelihood

Question 8

Which measure penalizes large forecast errors more heavily than MAE?

A. RMSE

B. Median demand

C. On-time delivery

D. Cost per kilogram

Correct Answer: A. RMSE

Question 9

What is data leakage in predictive freight modelling?

A. Loss of data during system maintenance

B. Use of information that was unavailable when the prediction would have been made

C. Duplicate shipment records only

D. Encryption of customer data

Correct Answer: B. Use of information that was unavailable when the prediction would have been made

Question 10

Which is the best governance practice for an AI-assisted customs or freight decision?

A. Allow autonomous approval without review

B. Suppress model assumptions from users

C. Require validation, source controls, and accountable human authorization

D. Use only the cheapest recommendation

Correct Answer: C. Require validation, source controls, and accountable human authorization