Multi-Criteria, Portfolio and Group Decision Models

Introduction and context

Many managerial decisions cannot be reduced to a single objective. A bank selecting a technology platform may consider cost, cybersecurity, implementation time, regulatory compliance, customer experience, and strategic flexibility. An investment committee may balance expected return against risk, liquidity, environmental exposure, and concentration. A public organization may compare projects according to social value, equity, fiscal cost, and political feasibility.

These are multi-criteria decision-making (MCDM) problems. They differ from ordinary optimization because alternatives are often evaluated against heterogeneous criteria, some quantitative and some qualitative, measured in different units and valued differently by stakeholders. A sound decision model therefore does more than produce a ranking. It makes preferences explicit, exposes trade-offs, identifies sensitivity to assumptions, and provides an auditable basis for discussion.

This chapter develops five connected approaches:

• weighted-scoring models;
• the analytic hierarchy process (AHP);
• outranking methods;
• Pareto efficiency and trade-off analysis; and
• portfolio optimization and group decision models.

The central principle is that a model should support judgment rather than conceal it. Mathematical sophistication cannot compensate for poorly defined objectives, unreliable data, double-counted criteria, or unexamined power relationships among stakeholders.

Learning objectives

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

1. formulate a multi-criteria decision problem with alternatives, criteria, measures, and constraints;
2. construct and interpret a weighted-scoring model;
3. apply AHP and assess the consistency of pairwise judgments;
4. distinguish compensatory scoring from outranking and Pareto approaches;
5. formulate a portfolio optimization model involving financial and nonfinancial objectives;
6. analyze sensitivity, robustness, and stakeholder disagreement;
7. design a transparent group decision process with appropriate governance; and
8. evaluate the ethical and practical limitations of decision models.

1. Formulating a multi-criteria decision problem

A decision model begins with five elements:

A criterion must be decision-relevant and sufficiently distinct from other criteria. If “system uptime” and “reliability” measure essentially the same construct, including both may double-count one benefit. Criteria should also be directionally defined: higher may be better for return and lower may be better for cost or risk.

A useful distinction is between objectives, which express what the decision maker wants, and attributes, which measure how well an alternative achieves an objective. For example, “improve resilience” is an objective, while “mean time to recovery” and “number of critical suppliers” are attributes.

Measurement and normalization

Because criteria commonly use different units, scores must be placed on a common scale. A simple linear benefit normalization is:

where x_ij is the observed performance of alternative i on criterion j, and z_ij lies between 0 and 1. This formula assumes that larger values are preferred and that the observed minimum and maximum are meaningful reference points.

For a cost criterion, lower is preferred:

If a target value is preferred rather than an extreme, a target-based transformation may be more appropriate. For example, a deviation score can be defined as:

where t_j is the desired target and d_j is the maximum acceptable deviation. Values should be bounded at zero when deviations exceed the acceptable range. Normalization is not neutral: changing the reference points can change the ranking.

2. Weighted-scoring models

The weighted-sum model is the most transparent compensatory MCDM method. Its general form is:

where V_i is the overall value of alternative i, w_j is the weight of criterion j, z_ij is the normalized score, and m is the number of criteria. Usually, w_j ≥ 0 and:

The model is compensatory: poor performance on one criterion can be offset by strong performance on another. This is appropriate only when such trade-offs are acceptable. A regulatory minimum, safety threshold, or non-negotiable ethical requirement should be represented as a constraint rather than merely as a weighted criterion.

Worked example: selecting a data platform

A company evaluates three platforms using cost, security, implementation speed, and strategic flexibility. The scores have already been normalized from 0 to 1. Cost and implementation speed are cost/time criteria, so higher normalized scores indicate better performance.

The total scores are:

Platform C ranks first, followed by B and A. However, this result is meaningful only if the normalization, weights, and compensatory assumptions are defensible. If security below 0.75 is unacceptable, Platform A should be eliminated before scoring rather than allowed to compete through low cost and high speed.

Sensitivity and robustness

A ranking should not be treated as definitive when small changes in assumptions reverse it. One-way sensitivity analysis varies one weight while holding others constant. Because weights must still sum to one, the changed weight can be offset proportionally across the remaining criteria or the analysis can use a structured scenario set.

For two alternatives, the break-even weight for criterion k can be found by setting their scores equal:

Solving for w_k identifies the weight at which the preferred alternative changes. Scenario analysis may compare, for example, a cost-focused, risk-focused, and balanced preference profile. Monte Carlo sensitivity analysis can assign probability distributions to uncertain scores and weights, then estimate the proportion of simulations in which each alternative ranks first. This produces a robustness perspective rather than a false point estimate.

3. The analytic hierarchy process

The analytic hierarchy process (AHP) decomposes a decision into a hierarchy of goal, criteria, subcriteria, and alternatives. Decision makers compare elements pairwise using a verbal or numerical scale, commonly from 1, meaning equal importance, to 9, meaning extreme importance. Reciprocal values represent the reverse comparison.

For a pairwise comparison matrix A, the element a_jk expresses the importance of criterion j relative to criterion k. The matrix satisfies:

A priority vector can be estimated by the normalized geometric mean method:

This method is convenient and often close to the principal-eigenvector method. The eigenvector approach obtains w from:

where λ_max is the largest eigenvalue. The weights are normalized so that they sum to one.

AHP also assesses judgment consistency. The consistency index is:

The consistency ratio is:

where RI is the average random index for matrices of size m. A commonly used practical guideline is that CR ≤ 0.10 indicates acceptable consistency, although this is not an inviolable statistical law. A high ratio signals that judgments should be reviewed, not that the decision maker has made a mathematical error.

AHP is valuable when stakeholders find direct numerical weighting difficult. Its limitations include many comparisons as the number of elements grows, sensitivity to hierarchy design, possible rank reversal under some procedures, and the assumption that criteria can be represented through ratio-scale judgments. AHP should therefore be used as a structured elicitation method, not as a machine that discovers objectively correct preferences.

4. Outranking methods

Weighted sums assume that performance differences can be traded across criteria using a common additive scale. Outranking methods take a different view. Alternative A outranks alternative B when the evidence favoring A is sufficiently strong and there is no criterion on which B is unacceptably better.

The two major families are ELECTRE and PROMETHEE. Their implementations vary, but they commonly use:

• criterion-specific preference thresholds;
• indifference thresholds for practically negligible differences;
• veto thresholds for severe disadvantages;
• concordance, or the extent of support for A over B; and
• discordance, or evidence that strongly contradicts the claim that A is at least as good as B.

For a simplified concordance calculation with normalized criterion weights:

where J(A,B) is the set of criteria on which A is at least as good as B according to the selected preference rule. A outranking relation may require C(A,B) to exceed a threshold while no veto condition is triggered.

Outranking is useful when criteria are incommensurable, when “small” and “large” differences should be treated differently, or when a major weakness cannot be compensated by many minor strengths. It can produce incomparability: neither A nor B outranks the other. This is not a defect; it may accurately represent conflicting evidence. The challenge is that thresholds and procedures require careful interpretation and can be less intuitive to nontechnical participants than a weighted score.

5. Pareto efficiency and transparent trade-offs

An alternative A Pareto-dominates alternative B if A is at least as good as B on every criterion and strictly better on at least one, assuming all criteria are oriented so that larger is better. Formally, A dominates B if:

A solution is Pareto-efficient if no feasible alternative dominates it. The set of efficient solutions is the Pareto frontier. A frontier does not identify one best solution; it identifies choices for which improving one objective requires sacrificing at least one other objective.

Pareto analysis is particularly important in sustainability, investment, and public policy. It prevents a decision maker from selecting an alternative that is worse in every relevant respect. It also provides a constructive basis for negotiation: stakeholders can focus on the efficient set and then discuss the values that determine the final choice.

A frontier may be generated by the ε-constraint method. Select one objective for optimization and impose minimum acceptable levels on others:

Here x represents decisions, f_r(x) represents objective functions, ε_r are acceptable thresholds, and X is the feasible region. Varying the ε values traces different efficient solutions. The method is useful when decision makers can state minimum acceptable performance more easily than precise weights.

6. Portfolio optimization with financial and nonfinancial objectives

Portfolio decisions allocate limited resources among projects, investments, suppliers, or strategic initiatives. Unlike a single-project ranking, portfolio selection must account for interactions, dependencies, diversification, and resource limits.

Let x_i be the proportion of capital allocated to asset or project i, with expected return μ_i. The expected portfolio return is:

where x and μ are vectors and n is the number of assets. If Σ is the covariance matrix of returns, portfolio variance is:

and portfolio standard deviation is σ_p = √(σ_p²). The covariance matrix must be symmetric and positive semidefinite for a valid variance model. Returns must be measured over a consistent period, such as annual expected returns and annualized covariance estimates.

A mean–variance model can minimize risk for a required return:

Here R^* is the target expected return, 1^Tx=1 imposes full investment, and x_i≥0 prohibits short selling. Real portfolios may add maximum allocation limits, turnover costs, liquidity constraints, minimum holdings, currency exposures, and regulatory restrictions.

Nonfinancial objectives can be integrated in several ways. A portfolio-level weighted objective might be:

where S(x) is a social or sustainability score, E(x) is an exposure or harm measure, and α,β,γ,δ express preferences. The terms must be scaled so that their units are comparable; otherwise the coefficients have no meaningful interpretation. An alternative is to use hard constraints, such as carbon intensity no greater than a specified limit, or to generate a Pareto frontier between return, risk, and impact.

Portfolio diversification and correlation

Diversification benefits arise because the variance of a two-asset portfolio is:

where ρ₁₂ is the correlation between returns. When correlation is below one, combining assets can reduce risk relative to a concentrated holding. Historical correlations are estimates, not permanent facts. Stress testing should examine adverse regimes in which correlations rise and liquidity falls.

7. Group decision and stakeholder models

Group decisions introduce information diversity, preference conflict, strategic behavior, and power asymmetry. A technically valid aggregation can still be procedurally unfair if some stakeholders cannot participate effectively or if criteria were chosen before affected groups were consulted.

A defensible group process normally includes:

1. Framing: agree on the decision question, scope, and authority.
2. Criteria development: identify objectives and remove overlaps.
3. Individual elicitation: collect judgments independently before discussion to reduce anchoring.
4. Facilitated deliberation: explain disagreements and revise unclear judgments.
5. Aggregation: combine judgments using a declared rule.
6. Sensitivity analysis: show how results change across stakeholder priorities.
7. Decision and documentation: record rationale, dissent, assumptions, and review dates.

For quantitative criteria, individual weights can be aggregated by an arithmetic mean:

where K is the number of participants. A weighted mean may be appropriate when participants have explicitly agreed that different institutional roles carry different decision authority. A geometric mean is often used to aggregate pairwise comparison judgments in group AHP:

The aggregation rule should be selected before results are known. Majority voting is appropriate for some discrete choices but can violate transitivity when preferences cycle. For example, one group may prefer A to B, another B to C, and a third C to A. Deliberation, ranked voting, or explicit negotiation may be needed rather than assuming that a simple vote reveals a stable collective preference.

Stakeholder analysis should distinguish preference disagreement, factual disagreement, and value disagreement. Better data can address factual disagreement; it cannot resolve a genuine conflict between, for example, financial efficiency and distributional equity. The model should make that conflict visible.

8. Illustrative case: selecting a sustainable logistics portfolio

Background

The fictional Northstar Retail Group must select a portfolio of logistics projects for a two-year investment program. Candidate projects include warehouse automation, electric delivery vehicles, route-optimization software, and supplier-consolidation initiatives. The board wants financial value, while operations emphasizes resilience and employees emphasize emissions and safety.

Problem

The organization has a budget of 10 million currency units and cannot fund every project. It must choose a portfolio rather than simply rank individual projects because projects compete for capital and may have dependencies.

Analysis

The decision team defines four criteria: discounted net present value (NPV), resilience improvement, emissions reduction, and implementation risk. NPV is measured in millions; the other criteria are scored by validated assessment scales. Projects must pass a minimum legal and safety screen before entering the optimization model.

A preliminary weighted score uses weights of 0.40 for NPV, 0.25 for resilience, 0.20 for emissions, and 0.15 for inverse risk. The team first normalizes each criterion using agreed reference ranges. It then identifies projects that are Pareto-dominated and removes them from further consideration. This prevents a dominated project from consuming meeting time.

The portfolio model uses binary variables y_i, where y_i=1 if project i is selected. A basic formulation is:

Here v_i is the composite project value, c_i is cost in millions, e_i is emissions reduction, and E_min is the required portfolio minimum. If automation requires route-optimization software, the dependency is represented by y_automation ≤ y_software.

The analysts generate three portfolios: a value-maximizing portfolio, an emissions-focused portfolio, and a balanced portfolio. The value-maximizing portfolio produces the highest NPV but narrowly misses the board’s desired emissions ambition. The emissions-focused portfolio meets the ambition but has greater execution risk. The balanced portfolio has slightly lower NPV than the financial maximum, materially lower emissions, and better resilience.

Approach and outcome

The committee does not choose the highest composite score automatically. It reviews the Pareto-efficient portfolios, examines the break-even weight at which the emissions-focused portfolio becomes preferred, and stress-tests NPV under lower demand and higher energy prices. It selects the balanced portfolio, adds a staged funding gate after six months, and establishes indicators for realized savings, delivery emissions, service disruptions, and employee safety.

Learning points

The case illustrates four principles. First, project ranking is not portfolio optimization. Second, constraints can express non-negotiable policy requirements more clearly than weights. Third, a Pareto frontier enables informed value discussion. Fourth, implementation governance is part of the decision: a portfolio should include monitoring, stop rules, and reassessment when assumptions change.

9. Governance, ethics, and implementation

MCDM systems can create an appearance of objectivity while embedding unexamined values. Governance should address at least the following issues:

• Transparency: publish criteria, definitions, weights, thresholds, data sources, and exclusions.
• Auditability: preserve versions of data, models, judgments, and approvals.
• Fairness: assess whether criteria disadvantage protected or vulnerable groups.
• Participation: include affected stakeholders, not only those with formal authority.
• Conflicts of interest: require declarations and independent review where incentives are material.
• Privacy: minimize personal data and control access to sensitive stakeholder information.
• Explainability: provide decision-relevant reasons rather than only a final score.
• Accountability: assign a human decision owner; a model should not become a mechanism for evading responsibility.

AI can assist with data extraction, scenario generation, inconsistency detection, and sensitivity analysis. It should not silently infer stakeholder values, fabricate evidence, or make high-impact decisions without review. Model outputs require validation for data drift, proxy discrimination, and instability. In regulated settings, documentation and human oversight may be mandatory.

10. Limitations and best practices

No MCDM method is universally superior. Weighted scoring is accessible but can hide compensation and scale effects. AHP structures judgment but may require many comparisons and depends on hierarchy design. Outranking represents vetoes and incomparability but can be procedurally complex. Pareto analysis avoids premature aggregation but leaves the final choice unresolved. Portfolio optimization is rigorous about constraints and interactions but is sensitive to estimation error and often requires assumptions that are difficult to communicate.

Best practice is therefore methodological pluralism. A robust workflow is:

1. screen for legal, safety, and feasibility requirements;
2. define objectives and remove overlapping criteria;
3. use reliable and comparable measures;
4. select a method whose assumptions match the decision;
5. involve stakeholders before weights are finalized;
6. calculate a base case and at least one alternative method;
7. test sensitivity to weights, data, thresholds, and scenarios;
8. identify Pareto-efficient or robust alternatives;
9. document dissent and unresolved value conflicts; and
10. monitor outcomes and revise the model when evidence changes.

A decision model is successful when it improves the quality of reasoning, not merely when it produces a numerically precise ranking.

Chapter Summary

Multi-criteria decision models support choices involving competing financial, operational, social, and strategic objectives. Weighted-sum models combine normalized performance scores using explicit weights, but their compensatory nature must be justified. AHP decomposes complex decisions and elicits priorities through pairwise comparisons, with consistency analysis helping identify problematic judgments. Outranking methods use preference, indifference, and veto concepts and can preserve incomparability where evidence does not support a complete ranking. Pareto analysis identifies efficient alternatives without imposing premature value trade-offs. Portfolio optimization extends analysis from individual choices to resource allocation, diversification, dependencies, and constraints. Group decision models require fair elicitation, transparent aggregation, deliberation, and documentation of disagreement. Across all methods, normalization, assumptions, sensitivity, governance, ethics, and post-decision monitoring are as important as the optimization calculation itself.

Key Terms

Multi-criteria decision-making: A structured approach for evaluating alternatives across multiple objectives.

Alternative: A feasible option being considered.

Criterion: A dimension used to evaluate alternatives.

Attribute: An observable measure of performance on a criterion.

Normalization: Transformation of measurements into a common scale.

Weighted-sum model: A compensatory model that multiplies criterion scores by weights and adds them.

Compensation: The ability of strong performance on one criterion to offset weak performance on another.

Analytic hierarchy process: A hierarchical pairwise-comparison method for deriving priorities.

Consistency ratio: An indicator of inconsistency in AHP judgments.

Outranking: A relation in which one alternative is sufficiently supported as at least as good as another.

Concordance: Evidence supporting an outranking claim.

Veto threshold: A disadvantage large enough to block an outranking relation.

Pareto dominance: A condition in which one alternative is no worse on every criterion and better on at least one.

Pareto efficiency: A state in which no feasible alternative dominates the selected alternative.

Portfolio optimization: Allocation of resources across multiple projects or assets under objectives and constraints.

Covariance matrix: A matrix describing variances and pairwise covariances among returns.

Diversification: Reduction of portfolio-specific risk through combining imperfectly correlated holdings.

Sensitivity analysis: Examination of how results change when assumptions vary.

Group decision-making: A process for combining judgments and preferences from multiple stakeholders.

Decision governance: Rules for accountability, transparency, participation, review, and control of decision models.

Review Questions

Short Answer Questions

1. Why must criteria be normalized before applying a weighted-sum model?
2. What is the difference between a criterion and a constraint?
3. What does a high AHP consistency ratio indicate?
4. Define Pareto dominance.
5. Why can portfolio optimization produce different results from ranking projects individually?

Long Answer Questions

1. Explain the assumptions and limitations of the weighted-sum model, using a managerial example.
2. Compare AHP and outranking methods in terms of preference elicitation, compensation, and interpretability.
3. Explain how a Pareto frontier can support negotiation among stakeholders with conflicting objectives.
4. Discuss the role of covariance and correlation in portfolio diversification and identify important estimation risks.
5. Design a governance process for a high-impact group decision model involving financial and social criteria.

Analytical / Application Questions

1. A firm evaluates two suppliers on cost, resilience, and emissions. Construct a normalized weighted-score model, calculate the rankings under two different weight profiles, and identify whether the result is robust.
2. Develop a portfolio model for three projects with budget, dependency, and minimum sustainability requirements. State the decision variables, objective, and constraints, and explain how you would test sensitivity.
3. A committee’s financial, operations, and community representatives disagree substantially on weights. Propose an aggregation and deliberation process, and explain how you would distinguish factual disagreement from value disagreement.

Multiple Choice Questions with Answers

Question 1

Which condition must normally hold for a standard weighted-sum model with normalized scores?

A. All criteria must be measured in currency.

B. The weights must be nonnegative and sum to one.

C. Every alternative must have identical performance.

D. All criteria must be maximized without transformation.

Correct Answer: B. The weights must be nonnegative and sum to one.

Question 2

In the data-platform example, which platform received the highest weighted score?

A. Platform A

B. Platform B

C. Platform C

D. All platforms tied

Correct Answer: C. Platform C

Question 3

What does a high AHP consistency ratio primarily signal?

A. The alternatives are infeasible.

B. The pairwise judgments may require review.

C. The weights must all be equal.

D. The decision has no Pareto solution.

Correct Answer: B. The pairwise judgments may require review.

Question 4

An alternative Pareto-dominates another when it is:

A. Better on every criterion, including at least one strictly better result.

B. Worse on every criterion but less expensive.

C. Preferred by a majority of stakeholders.

D. The alternative with the highest weighted score.

Correct Answer: A. Better on every criterion, including at least one strictly better result.

Question 5

Which feature is most characteristic of an outranking method?

A. It always produces a complete ranking.

B. It requires all criteria to use the same physical units.

C. It can use veto thresholds and permit incomparability.

D. It excludes stakeholder judgments.

Correct Answer: C. It can use veto thresholds and permit incomparability.

Question 6

If the covariance between two assets is low relative to their individual variances, combining them may:

A. Increase diversification benefits.

B. Guarantee a positive return.

C. Eliminate all market risk.

D. Make expected return undefined.

Correct Answer: A. Increase diversification benefits.

Question 7

In a portfolio model, the constraint ∑x_i = 1 commonly represents:

A. Full allocation of the available portfolio capital.

B. A requirement that all returns equal one.

C. A prohibition on diversification.

D. A minimum emissions threshold.

Correct Answer: A. Full allocation of the available portfolio capital.

Question 8

Which is the best treatment for a non-negotiable safety minimum?

A. Represent it only through a small weight.

B. Represent it as a feasibility constraint or screening rule.

C. Ignore it until after ranking.

D. Replace it with the cheapest criterion.

Correct Answer: B. Represent it as a feasibility constraint or screening rule.

Question 9

What is the purpose of sensitivity analysis in MCDM?

A. To guarantee that the preferred alternative is correct.

B. To determine whether conclusions change when assumptions vary.

C. To remove all stakeholder preferences.

D. To convert qualitative criteria into legal requirements.

Correct Answer: B. To determine whether conclusions change when assumptions vary.

Question 10

Which practice most improves accountability in a group decision model?

A. Keeping weights confidential from participants.

B. Changing the aggregation rule after seeing the result.

C. Documenting criteria, weights, data sources, dissent, and approvals.

D. Allowing the model to make the final decision without review.

Correct Answer: C. Documenting criteria, weights, data sources, dissent, and approvals.