Logistics Analytics and Decision Support

Introduction

Modern logistics systems generate extensive data from enterprise resource planning systems, transportation-management systems, warehouse-management systems, telematics, barcode scans, GPS devices, electronic proof of delivery, customer orders, weather services and external market databases. Data alone, however, does not create operational value. Value arises when data is converted into reliable information, information into insight, and insight into coordinated decisions.

Logistics analytics is the systematic use of data, statistical methods, optimization models, simulation and managerial judgement to improve the movement and storage of goods. It supports decisions at three levels:

• Strategic: network design, facility location, fleet ownership and modal strategy.
• Tactical: capacity procurement, inventory positioning, carrier allocation and seasonal planning.
• Operational: shipment consolidation, route selection, dispatching, exception management and delivery control.

This chapter examines descriptive, predictive and prescriptive analytics in multimodal freight systems. It emphasizes not only model construction but also data quality, assumptions, managerial interpretation, implementation and governance. A mathematically optimal solution is not necessarily a good business decision if its data is unreliable, its constraints are incomplete or its recommendations cannot be executed.

Learning Objectives

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

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Construct and interpret logistics performance measures and dashboards.
3. Apply appropriate forecasting methods for freight demand.
4. Explain capacity-planning, routing and network-optimization models.
5. Use sensitivity analysis and simulation to evaluate uncertainty.
6. Assess the role and limitations of artificial intelligence in logistics decisions.
7. Design governance practices for trustworthy logistics analytics.

1. The Analytics Hierarchy in Logistics

A useful analytics hierarchy moves from understanding the past to recommending action.

These categories are related but not interchangeable. A dashboard showing a 12% late-delivery rate is descriptive. Identifying port congestion and incomplete documentation as major causes is diagnostic. Estimating next week’s late-delivery probability is predictive. Selecting alternative ports, modes and departure times to minimize expected cost is prescriptive.

The analytics lifecycle normally contains six stages: define the decision, acquire and prepare data, explore and model, validate results, deploy recommendations, and monitor outcomes. Beginning with a sophisticated algorithm before defining the decision often produces technically impressive but operationally irrelevant work.

Decision framing

A logistics decision should specify:

• the decision-maker and decision horizon;
• controllable variables, such as shipment allocation or departure time;
• objectives, such as cost, service, resilience or emissions;
• constraints, such as vehicle capacity, delivery windows and regulatory rules;
• uncertainty, including demand, travel time and equipment availability;
• the performance measure used to judge the decision.

A single objective can be misleading. For example, minimizing transport cost may increase late deliveries, carbon emissions or exposure to a single carrier. Practical models therefore use multiple objectives or explicitly assign penalties and weights to competing outcomes.

2. Data Architecture and Data Quality

Important logistics data entities include orders, shipment legs, locations, products, vehicles, containers, carriers, rates, schedules, inventory positions and events. A multimodal shipment should be represented as a sequence of linked legs rather than as one undifferentiated movement. Each leg may have its own mode, lead time, cost, capacity, emissions factor and operational risk.

A basic analytical data model distinguishes:

• Facts: quantities, costs, times, distances, scans and service outcomes.
• Dimensions: date, origin, destination, product, carrier, mode and customer.
• Master data: location codes, calendars, vehicle types, packaging units and service rules.
• Event data: booking, pickup, departure, transshipment, arrival, customs release and delivery.

Data quality has several dimensions:

Time definitions require special care. “On time” may mean arrival before the customer promise, before the requested delivery date, or within a contractual tolerance. Analysts should document event definitions, time zones, cut-off rules, cancellations and partial deliveries.

For a KPI based on a sample, the arithmetic mean transit time is:

Here, t_i is the transit time of shipment i, n is the number of shipments, and \bar{t} has the same time unit as t_i. The mean is appropriate when extreme values are meaningful and data is sufficiently reliable. The median is often more representative when delays create a strongly right-skewed distribution.

A useful service measure is the on-time delivery rate:

N_on-time is the number of eligible shipments meeting the defined promise and N_eligible is the total eligible population. The analyst must state whether partial deliveries, customer-requested delays and cancelled orders are included.

3. Descriptive and Diagnostic Analytics

Descriptive analytics converts operational records into meaningful views. Common logistics measures include:

• shipment volume by origin, destination, customer and mode;
• cost per shipment, kilogram, pallet, kilometre or tonne-kilometre;
• cube and weight utilization;
• dwell time at terminals;
• schedule adherence;
• order cycle time;
• claims, damage and loss rates;
• empty kilometres and asset utilization;
• emissions per shipment or tonne-kilometre.

A cost-per-weight measure can be expressed as:

where C_total is the relevant transport cost in currency and W is transported weight in kilograms or tonnes. It is meaningful only when the cost boundary, weight basis and shipment population are consistent.

Dashboards should be designed around decisions rather than around the maximum number of charts. An effective logistics dashboard includes a small number of headline KPIs, trend comparisons, exception thresholds, drill-down capability and clear ownership. A red indicator should lead to an action, not merely signal failure.

Diagnostic analytics investigates causation without claiming that correlation alone proves causality. Useful methods include Pareto analysis, process mapping, cohort comparison, control charts, regression, root-cause analysis and delay decomposition. Comparisons should control for relevant factors: lane, shipment size, carrier, season, weather, service level and mode. Comparing all carriers without controlling for lane mix can unfairly penalize carriers serving difficult routes.

Control charts can distinguish routine variation from unusual variation. However, their assumptions must be considered: observations should be measured consistently, sampling should be rational, and serial dependence or changing operating conditions should not be ignored. A statistically unusual event is not automatically operationally important, and a stable process can still be unacceptable if its average performance misses the customer requirement.

4. Predictive Analytics and Freight Forecasting

Forecasting estimates future demand, volume, transit time or disruption probability. Forecasts are not certainties; they are conditional statements based on historical patterns, explanatory variables and assumptions.

4.1 Forecast components

Freight demand may contain:

• level: the underlying average;
• trend: systematic growth or decline;
• seasonality: repeating calendar effects;
• cycles: longer fluctuations linked to economic conditions;
• irregular variation: random or exceptional events;
• causal effects: prices, promotions, weather, holidays and industrial activity.

The planning horizon matters. A same-day dispatch forecast can use current order arrivals and GPS data, while a five-year network forecast requires scenarios for market growth, regulation and infrastructure.

4.2 Basic forecasting methods

The moving-average forecast for the next period using the previous k observations is:

A_t is actual demand in period t, F_t+1 is the forecast for the next period, and k is the window length. This method is simple and smooths random variation, but it lags when demand changes rapidly and gives equal weight to all observations in the window.

Exponential smoothing gives greater weight to recent observations:

Here, α is the smoothing parameter, A_t is actual demand and F_t is the forecast for period t. A high α reacts quickly but may follow noise; a low value is smoother but slower to respond.

For demand with trend and seasonality, methods such as Holt’s trend method, Holt–Winters exponential smoothing, seasonal decomposition and ARIMA may be appropriate. Regression or machine-learning models can incorporate explanatory variables such as promotions, commodity output, weather and economic indicators. Model choice should follow the demand process, data availability and decision use—not fashion.

4.3 Forecast accuracy

Common error measures include:

MAE is expressed in the original demand unit and treats errors linearly. RMSE also uses the original unit but penalizes large errors more heavily. Percentage measures can be problematic when actual demand is zero or near zero. Forecasts should be evaluated using rolling or time-ordered validation, not random shuffling that allows future information to leak into the training set.

The forecast should be accompanied by prediction intervals or scenarios. A point estimate of 1,000 pallets is less useful than a forecast of 1,000 pallets with a credible range and an explanation of its drivers. Bias is especially important: repeated under-forecasting causes chronic capacity shortages, while repeated over-forecasting creates costly unused capacity.

5. Capacity Planning

Capacity planning translates expected freight into required transport, terminal and warehouse resources. It must account for volume, weight, cube, time windows, service reliability, repositioning and variability.

Suppose daily demand is forecast as D tonnes, each vehicle carries q tonnes, and the planned utilization fraction is u. The minimum loaded vehicle-equivalents required, ignoring other constraints, are:

The ceiling operator rounds upward because a fractional vehicle cannot be dispatched. This is a simplified calculation. Actual requirements may be higher because of route incompatibility, driver hours, delivery windows, empty returns, maintenance and demand variability.

Capacity can be secured through owned fleets, dedicated contracts, spot markets, rail slots, vessel bookings or shared-user facilities. Each alternative differs in fixed cost, flexibility, service control and risk. Capacity planning should therefore compare total expected cost, not only nominal rate.

For uncertain demand, a robust plan may reserve flexible capacity. The value of flexibility can be assessed by comparing the expected cost of shortage, excess capacity and option procurement. Scenario analysis is often more credible than a single deterministic forecast: for example, low, base and high demand combined with normal and disrupted transit conditions.

6. Route Selection and Multimodal Choice

Route selection is a constrained decision involving cost, transit time, reliability, capacity, handling risk, emissions and regulatory feasibility. In multimodal transportation, a route is a chain of legs connected at transfer nodes. The fastest individual leg may not produce the fastest end-to-end movement if transfer dwell and customs time are substantial.

For a candidate route r, a generalized cost can be represented as:

C_r is monetary transport and handling cost, T_r is transit time, v is the value assigned per unit of time, P_r is the probability or expected magnitude of service failure as defined by the model, p_L is its penalty, E_r is emissions, and e is the emissions valuation. Units must be compatible: every term must be expressed in the same generalized monetary unit. The expression is a decision model, not a universal law; weights should reflect organizational priorities and stakeholder requirements.

Shortest-path algorithms are useful when link costs are known and additive. Dijkstra’s algorithm requires non-negative arc costs and finds shortest paths from a source in a network with fixed costs. Time-dependent networks require costs that vary by departure time, while vehicle-routing problems add vehicle capacity, customer windows, depot constraints and sometimes pickup-and-delivery precedence. These extensions are computationally harder and often use mixed-integer programming, heuristics or metaheuristics.

A route recommendation should show alternatives and trade-offs. For example, rail may have lower cost and emissions but longer and less predictable transit; air may provide speed at high monetary and environmental cost; road may offer door-to-door flexibility but expose the shipment to congestion and driver constraints.

7. Network Optimization

Network optimization determines how flows move through origins, facilities, transshipment points and destinations. A simplified minimum-cost flow model uses:

• x_ij: quantity shipped from node i to node j;
• c_ij: unit cost on arc (i,j);
• s_i: supply at origin i;
• d_j: demand at destination j;
• u_ij: capacity of arc (i,j).

The objective is:

subject to supply, demand and capacity constraints:

The model assumes costs are known, additive and appropriately represented by the selected arcs. If supply equals demand and all demand must be met, the inequalities may be written as equalities. In practice, shortage variables, fixed facility costs, service penalties, carbon limits, mode restrictions and integer decisions may be added.

Facility-location models trade fixed facility cost against transportation cost and service. A typical objective includes the fixed cost of opening facilities, flow costs and penalties for unmet demand. The model can be extended with multiple periods, inventory balance, capacity expansion and disruption scenarios. Because strategic decisions are sensitive to assumptions, managers should inspect alternative scenarios rather than accept one optimal network as inevitable.

8. Prescriptive Analytics, Sensitivity and Simulation

Prescriptive analytics recommends actions under constraints. Optimization identifies a best solution for a specified model; simulation evaluates system behaviour under repeated or stochastic conditions. They are complementary.

Sensitivity analysis

Sensitivity analysis changes assumptions and observes the effect on decisions or outcomes. Relevant variables include demand, fuel price, handling cost, capacity, service penalty, transit time, carbon price and disruption probability.

A model is structurally sensitive when a small change changes the selected route, facility or carrier. A model is outcome sensitive when the selected decision remains the same but its cost or service result changes substantially. Both forms matter. A stable decision with rapidly worsening service under a disruption may require contingency planning.

Scenario analysis changes coherent groups of assumptions, such as “port congestion plus high demand,” rather than changing one variable arbitrarily. Tornado charts, break-even analysis and shadow-price interpretation can help managers understand which assumptions deserve better data.

Simulation

Discrete-event simulation represents events such as order arrival, picking, loading, departure, transfer, customs clearance and delivery. Monte Carlo simulation repeatedly samples uncertain demand, travel time or processing time to estimate distributions of cost, delay and resource utilization.

A simulation is credible only if its process logic, input distributions and validation are credible. It should be calibrated against observed historical behaviour and tested with extreme but plausible cases. Random variation does not excuse poor process representation. Simulation results should be reported with confidence intervals or sufficient replication, not as a single falsely precise number.

Optimization under uncertainty can use stochastic programming, robust optimization or chance constraints. Robust optimization seeks solutions that remain acceptable across a defined uncertainty set. Stochastic programming uses probability distributions and expected or risk-sensitive objectives. These approaches differ from simply adding a large arbitrary safety margin.

9. Dashboards and Decision Support Design

A decision-support system should connect data, model, recommendation and action. A practical architecture includes data sources, an integration and quality layer, analytical models, a visualization layer, workflow integration and feedback from actual outcomes.

Dashboard design principles include:

1. Define the decision and audience before selecting visualizations.
2. Display definitions, units, time period and data freshness.
3. Separate leading indicators, such as booking backlog, from lagging indicators, such as delivered-on-time rate.
4. Use thresholds linked to escalation rules.
5. Permit drill-down from network performance to shipment and event level.
6. Show uncertainty and confidence rather than hiding it.
7. Preserve an audit trail of data, model version and user overrides.

A dashboard can create harmful behaviour when metrics are poorly designed. Measuring only vehicle utilization may encourage consolidation that violates delivery commitments. Measuring only cost per shipment may encourage shipment splitting or deferred maintenance. Balanced scorecards should include cost, service, asset productivity, resilience, safety, compliance and environmental performance.

10. Artificial Intelligence and Advanced Analytics

Machine learning can support demand forecasting, estimated-time-of-arrival prediction, anomaly detection, document extraction, dynamic pricing and predictive maintenance. Graph-based learning can represent network relationships; reinforcement learning can investigate sequential dispatch decisions; natural-language systems can assist planners in querying operational data.

These tools do not eliminate the need for operations expertise. Risks include training-data bias, concept drift, data leakage, opaque recommendations, adversarial or corrupted data, automation bias and weak out-of-distribution performance. A model trained during stable operations may fail during strikes, pandemics, severe weather or regulatory change.

Good practice includes time-based validation, champion–challenger testing, drift monitoring, human approval for high-impact decisions, explainability appropriate to the decision, and controlled rollback. Generative AI should not be allowed to invent shipment status, compliance evidence or customer commitments. Its output must be grounded in authoritative systems and subject to access control.

11. Governance, Ethics and Regulation

Logistics analytics involves personal, commercial and operationally sensitive information. Telematics may reveal driver behaviour and location; customer data may reveal commercial relationships; shipment information may involve controlled goods. Governance should address lawful collection, purpose limitation, retention, access rights, security, encryption, data minimization and incident response in accordance with applicable privacy and sector regulations.

Important governance controls include:

• data ownership and stewardship;
• documented definitions and lineage;
• role-based access;
• model documentation and version control;
• validation and approval thresholds;
• monitoring for discriminatory or unsafe outcomes;
• segregation of duties for model development and deployment;
• human escalation for exceptional decisions;
• records supporting audit and regulatory review.

Ethically, optimization should not externalize risk onto drivers, communities or vulnerable suppliers. A route that minimizes cost by creating unsafe schedules is unacceptable. Environmental claims should use transparent boundaries and defensible emissions factors. Decisions affecting employment, surveillance or customer access require proportionality and accountability.

12. Illustrative Case: Northstar Consumer Goods

Background

Northstar Consumer Goods distributes packaged products from two plants to 18 regional markets. It uses road transport for short distances, rail for selected long-haul movements and coastal shipping for one corridor. Management observed rising premium-freight expenditure and inconsistent delivery performance.

Problem

The company’s dashboard reported an overall on-time delivery rate of 91%, but the measure excluded partial deliveries and used different promised-date definitions across business units. Forecasts were aggregated monthly, although carrier capacity and customer commitments were managed weekly. The network team proposed shifting more freight to rail solely because its average rate was lower.

Analysis

The analytics team first standardized event definitions and reconstructed shipment legs. It found that:

• demand on three lanes was highly seasonal;
• rail had lower line-haul cost but longer terminal dwell;
• premium road shipments were concentrated in weeks with forecast underestimation;
• one plant frequently missed the carrier booking cut-off;
• the apparent carrier performance difference was partly explained by lane mix.

The team developed a weekly forecast by lane using seasonal exponential smoothing and evaluated it through rolling-origin testing. It then built a mixed-integer allocation model with carrier capacity, delivery windows, mode eligibility and a penalty for late delivery. Scenarios represented normal operations, high demand and a coastal-port disruption.

Approach and solution

The recommended policy reserved a small amount of flexible road capacity, shifted predictable base volume to rail, introduced a booking-cut-off control at the plant, and used a disruption rule that rerouted selected coastal shipments through an alternative terminal. A dashboard displayed forecast bias, capacity reservation, late-risk shipments and the causes of exceptions.

The company did not use the lowest-cost solution as its sole policy. It selected a plan whose expected cost was slightly higher than the unconstrained minimum but which maintained service under the high-demand scenario and reduced dependence on one terminal. The model was implemented with planner override, and overrides were recorded for later model improvement.

Outcome and learning points

In the pilot period, premium-freight use declined, while the revised service measure improved because it reflected complete delivery performance. The most important improvement was not a particular algorithm; it was the alignment of data definitions, planning cadence, constraints and accountability.

The case illustrates five principles: clean the measurement system before optimizing; model the full multimodal journey; treat uncertainty explicitly; test solutions against disruption scenarios; and make recommendations executable through workflow and ownership.

13. Best-Practice Implementation Roadmap

Organizations beginning logistics analytics should proceed incrementally:

1. Identify high-value decisions and establish a common KPI dictionary.
2. Create trusted master data for locations, products, carriers and events.
3. Build descriptive dashboards with reconciled historical data.
4. Introduce diagnostic analysis for major cost and service failures.
5. Develop forecasts with time-based validation and bias monitoring.
6. Pilot prescriptive models in a contained lane, facility or planning cycle.
7. Integrate recommendations into operational workflows.
8. Establish model governance, user training and override procedures.
9. Measure realized benefits, not only model accuracy.
10. Retire or revise models when operating conditions change.

The maturity of analytics should be judged by decision quality and realized operational outcomes, not by the sophistication of software. A transparent linear model with good data may outperform an opaque machine-learning model that planners do not trust or cannot execute.

Challenges and Limitations

Logistics analytics faces persistent limitations. Historical data may encode past inefficiencies and therefore reproduce them. Forecast accuracy deteriorates at highly aggregated or highly disaggregated levels. Optimization models can become computationally difficult when they include time, inventory, uncertainty and integer decisions. Local improvements can damage network-wide performance. KPI incentives can generate gaming. Real-time decisions may be constrained by latency, missing data and conflicting systems.

Models also simplify reality. A carrier may accept a nominal allocation but lack practical driver or equipment availability. A route may be feasible mathematically but unacceptable because of political risk, community restrictions or customs uncertainty. Consequently, analytical recommendations require domain review, scenario testing and post-implementation measurement.

Chapter Summary

Logistics analytics supports decisions from strategic network design to real-time exception management. Descriptive analytics explains what happened; diagnostic analytics investigates why; predictive analytics estimates what may happen; and prescriptive analytics evaluates what should be done. Reliable analytics depends on consistent event definitions, quality master data, appropriate metrics and time-ordered validation.

Freight forecasting should represent level, trend, seasonality and uncertainty, while capacity planning must incorporate utilization, service windows and variability. Route and network models express cost, time, capacity, service and environmental trade-offs. Optimization and simulation are complementary: optimization selects actions within a model, whereas simulation tests performance under uncertainty. Sensitivity analysis reveals fragile assumptions and supports resilient decisions.

Dashboards are effective only when connected to decisions, thresholds and accountability. AI can improve prediction and exception handling but requires validation, monitoring, human oversight and governance. Ethical and regulatory responsibilities include privacy, security, transparency, safety and avoidance of harmful incentives. The central managerial lesson is that analytics is a socio-technical capability: value comes from combining data, models, operational knowledge and disciplined execution.

Key Terms

Analytics hierarchy: Progression from descriptive and diagnostic analysis to predictive and prescriptive decision support.

Capacity planning: Determining the resources required to meet expected demand and service requirements.

Diagnostic analytics: Analysis intended to identify causes of observed outcomes.

Descriptive analytics: Summarization and visualization of historical or current performance.

Exponential smoothing: Forecasting method that weights recent observations more heavily.

Forecast bias: Systematic tendency to over-forecast or under-forecast.

Generalized cost: Combined representation of monetary and non-monetary route attributes in a common decision metric.

Key performance indicator: Defined measure used to monitor operational or strategic performance.

Machine learning: Computational methods that learn patterns from data for prediction or classification.

Minimum-cost flow: Network model that allocates flows through capacitated arcs at minimum total cost.

Multimodal transportation: Movement using two or more transport modes within a coordinated journey.

Prescriptive analytics: Analysis that recommends actions subject to objectives and constraints.

Prediction interval: Range representing uncertainty around a future estimate.

Robust optimization: Optimization designed to remain acceptable across specified uncertainty conditions.

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

Simulation: Representation of system behaviour under repeated or stochastic conditions.

Time-series forecasting: Prediction using ordered observations over time.

Transit-time variability: Variation in elapsed movement time across shipments or conditions.

Data lineage: Record of a data element’s origin, transformations and use.

Model governance: Controls for developing, validating, deploying, monitoring and retiring analytical models.

Review Questions

Short Answer Questions

1. Distinguish descriptive, predictive and prescriptive analytics using one logistics example for each.
2. Why must “on-time delivery” be defined before it is reported as a KPI?
3. State two limitations of a simple moving-average forecast.
4. What is the difference between route cost and generalized cost?
5. Why are rolling-origin tests preferable to random train-test splits for time-series forecasting?

Long Answer Questions

1. Explain how data quality dimensions affect logistics decision support, using examples from multimodal shipment data.
2. Discuss the major considerations in designing a freight-capacity plan under uncertain demand.
3. Compare optimization and simulation as tools for logistics planning.
4. Explain how sensitivity analysis can reveal weaknesses in a network design recommendation.
5. Evaluate the governance and ethical requirements for using AI in transportation management.

Analytical / Application Questions

1. A carrier has a forecast of 840 tonnes per day. Each vehicle can carry 12 tonnes, and the planning utilization is 70%. Calculate the minimum vehicle-equivalents using V = ⌈D/(qu)⌉. Explain at least three reasons why the operational requirement could be higher.
2. A firm must choose between a low-cost rail route with high dwell-time variability and a faster road route with higher emissions. Design a generalized-cost comparison and identify the data required before making the recommendation.
3. Develop a scenario-based analytics plan for a distribution network facing high seasonal demand and a possible port closure. Specify the forecast, capacity, optimization, simulation and dashboard components.

Multiple Choice Questions with Answers

Question 1

Which analytics category recommends an action subject to objectives and constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

In the on-time delivery formula, the denominator represents:

A. All planned vehicles

B. Eligible shipments

C. Late shipments only

D. Total transport cost

Correct Answer: B. Eligible shipments

Question 3

A high value of the exponential-smoothing parameter α generally makes a forecast:

A. More responsive to recent observations

B. Independent of recent observations

C. Unable to use historical data

D. Always unbiased

Correct Answer: A. More responsive to recent observations

Question 4

Which measure penalizes large forecast errors more heavily than MAE?

A. Median error

B. RMSE

C. On-time rate

D. Utilization

Correct Answer: B. RMSE

Question 5

If daily demand is 840 tonnes, vehicle capacity is 12 tonnes, and utilization is 70%, what is the minimum vehicle-equivalent requirement?

A. 70

B. 90

C. 100

D. 120

Correct Answer: C. 100

Question 6

A key reason to reconstruct a multimodal shipment as linked legs is that:

A. Each leg can have different cost, time and capacity characteristics

B. All legs necessarily use the same carrier

C. Transfer time is always zero

D. Mode choice becomes irrelevant

Correct Answer: A. Each leg can have different cost, time and capacity characteristics

Question 7

Which condition is required by the standard Dijkstra shortest-path algorithm?

A. All arc costs must be negative

B. Arc costs must be non-negative

C. Every node must have equal capacity

D. Demand must be zero

Correct Answer: B. Arc costs must be non-negative

Question 8

What is the principal purpose of sensitivity analysis?

A. To remove all uncertainty from a model

B. To change historical records

C. To examine how outputs respond to changed assumptions

D. To guarantee the optimal solution is implemented

Correct Answer: C. To examine how outputs respond to changed assumptions

Question 9

Which practice helps prevent future information from contaminating time-series model evaluation?

A. Randomly shuffling all observations

B. Using time-ordered or rolling-origin validation

C. Deleting unusual observations automatically

D. Evaluating only the training data

Correct Answer: B. Using time-ordered or rolling-origin validation

Question 10

Why should an AI logistics recommendation include human oversight for high-impact decisions?

A. Models never require data

B. Operational conditions can change and models can fail outside training conditions

C. Human judgement makes validation unnecessary

D. Optimization cannot represent constraints

Correct Answer: B. Operational conditions can change and models can fail outside training conditions