GRA (Grey Relational Analysis)

GRA is a ranking method for when you hold a decision table filled with numbers and want the alternatives placed in a single order. Its output is a "grey relational grade" between 0 and 1 for every alternative, together with the rank that grade produces. The method comes out of "grey system theory," which Deng Julong proposed in 1982 and introduced in English in 1989; GRA is that theory's tool, adapted for multi-criteria decision-making. It does not generate weights; weights are supplied from outside.

An important clarification. The name "grey" has nothing to do with the type of the data. GRA works with crisp, fully known numerical data, just as TOPSIS or SAW do. The word "grey" comes from an analogy in Deng's system theory: a system is "white" when information about it is complete and clear, "black" when nothing is known, and "grey" when it is partly known. The "grey" in GRA's name does not mean the data is fuzzy or uncertain. DecisionMind holds a separate family of methods for fuzzy data (the fuzzy extensions), and GRA's "grey relational grade" must not be confused with the "membership degree" of those families; the two are different mathematical objects.

GRA's underlying idea is to ask "which alternative most resembles the reference." On each criterion, once normalised, the best possible value is taken as 1, and a reference sequence is built that holds this best value on every criterion, an alternative that does not actually exist. Every real alternative's absolute difference from this reference is measured on each criterion; small differences count as "close" to the reference (a high grey relation), large ones as "far" (a low grey relation). This is close to TOPSIS's logic of "distance to the ideal and the anti-ideal," but GRA looks at only a single reference, the ideal; it builds no anti-ideal counterpoint.

One consequence of this idea is a technical parameter called the "distinguishing coefficient" (ζ, usually 0.5). This coefficient sets how sharply or how gently the differences are reflected in the grey relational grade. GRA's philosophical consequence is compensatory, much as TOPSIS's is: a large difference on one criterion can be offset in the total by small differences on others, and there is no veto mechanism of the kind found in the ELECTRE family.

The method proceeds through five steps.

First, scale equalisation (normalisation). Criteria are expressed in different units. GRA places every column between 0 and 1 relative to its own minimum and maximum; direction is adjusted so that the largest value maps to 1 for a "higher is better" criterion, and the smallest value maps to 1 for a "lower is better" criterion.

Second, the reference sequence. Once normalisation is complete, the "best" value on every criterion is already 1; the reference sequence is therefore a hypothetical alternative, one that does not exist, holding 1 on every criterion.

Third, absolute differences. The absolute difference between each real alternative's normalised value and the reference sequence's value (1) is computed separately for every criterion.

Fourth, the grey relational coefficient. Each cell's absolute difference is converted into a coefficient between 0 and 1, using the smallest and largest absolute differences in the dataset together with a distinguishing coefficient (ζ, usually 0.5). A small difference pushes the coefficient towards 1; a large one pushes it towards 0.

Fifth, the grey relational grade and ranking. Each alternative's grey relational coefficients across the criteria are multiplied by the criterion weights and summed. This total is the "grey relational grade," showing how close the alternative sits to the reference overall. Alternatives are ranked from the highest grade to the lowest.

The formulas behind each step, the intermediate tables and the citation formats are given on the DecisionMind method page; this card carries no formulas.

The grey relational grade shows how closely an alternative tracks the hypothetical reference sequence relative to the other alternatives in this particular analysis, and nothing more. A grade of 0.77 does not mean "77 per cent good," nor is it a probability. It cannot be compared with a grade from a different analysis, because the reference sequence is built afresh in every analysis from that analysis's own criteria. A grade approaching 1 does not mean "perfect" but "closest to the best on every criterion within this set."

Thus instead of writing:

"GRA found the best alternative"

the report should read:

"With these weights and this distinguishing coefficient, the alternative closest to the reference is this one; the ranking is sensitive to the weight on these criteria"

GRA works with crisp data: one number per cell. The name "grey" refers to Deng's system-theory analogy for missing information, not to the data type. You need: alternatives in rows, criteria in columns, one number per cell with no empty cells; a direction for every criterion; criterion weights that sum to 1; and, optionally, a distinguishing coefficient (DecisionMind uses the conventional value of 0.5 if none is given).

GRA does not produce weights, it asks for them; they can be drawn from sources such as AHP, BWM, Entropy or CRITIC. DecisionMind holds twelve GRA family members alongside the base method, including fuzzy and intuitionistic fuzzy extensions; these extensions work with fuzzy data, whereas the base GRA does not. A minimum of two alternatives and two criteria is required; three to twelve criteria work comfortably. If every alternative takes the same value on a criterion, the normalisation denominator becomes zero; this must be checked before the data is entered.

GRA is a sound choice if your criteria can be measured numerically, your table is complete, your sample is small, or you do not want to make strong assumptions about the distribution of the data; the method rests on geometric closeness without assuming a parametric distribution. Its typical territory includes materials and supplier selection, optimisation of production process parameters, and environmental and energy assessments.

It should not be used where a criterion carries a threshold on which no compromise can be made; GRA's additive structure cannot preserve such a threshold on its own. Where your data is fuzzy, or is expert judgement described as "approximate," the relevant fuzzy extension should be used rather than the base method.

A numerical table, compensation accepted, a small sample → GRA

Same logic, but the data is fuzzy / intuitionistic fuzzy → the relevant GRA extension

No compromise allowed on one criterion, sub-threshold alternatives must be screened out → the ELECTRE family

The "closeness to the ideal" logic, but comparison against two endpoints (ideal and anti-ideal) is wanted → TOPSIS

Not a ranking but weights are needed → AHP, BWM, SWARA (subjective); Entropy, CRITIC (objective)

GRA's greatest strength is that it can work with a small sample and incomplete information: grey system theory was designed precisely for "partly known" systems, and the method requires no assumption about a parametric distribution. Its computational burden is light; working with a single reference sequence makes it simpler than TOPSIS's two-endpoint structure. As Kuo, Yang and Huang have shown, the method has found wide application in multi-criteria decision problems (Kuo, Yang and Huang, 2008). The distinguishing coefficient (ζ) allows the sensitivity of the result to the differences to be tuned, a flexibility not found in other ranking methods.

Its limitations stem from the same structure. First, the name "grey" is frequently misunderstood, and the method is often assumed to work with fuzzy data, whereas the base GRA requires crisp data. Second, the choice of the distinguishing coefficient (ζ) is a technical convention (0.5 by tradition); as Mahmoudi and colleagues have shown, the value of this coefficient can affect the ranking and should not be fixed arbitrarily (Mahmoudi et al., 2020). Third, the method assumes full compensation: a serious weakness on one criterion can be papered over by others.

Fourth, a constant-valued criterion column zeroes the normalisation denominator. Fifth, the quality of the weights lies outside the method itself; a flawless calculation built on poor weights still produces a poor ranking.

The most common mistake is confusing the name "grey" with data uncertainty and assuming GRA works with fuzzy or interval data; the base GRA requires crisp data, and the name comes only from Deng's system theory. A second mistake is leaving the distinguishing coefficient (ζ) at 0.5 without ever questioning it, ignoring how much it can affect the result. A third is marking criterion direction wrongly: mark a "lower is better" criterion as "higher is better" and the reference sequence is built incorrectly, reversing the ranking.

A fourth mistake is feeding a column into the analysis without noticing that every alternative takes the same value on that criterion. A fifth is reading the grey relational grade as a percentage or a probability, and comparing grades from different analyses.

The governing principle is this:

The grey relational grade is your alternative's crisp closeness to a hypothetical reference sequence; "grey" is not a data type but an analogy for missing information, and the grade cannot be compared with another analysis's grade.