ELECTRE II (ELimination Et Choix Traduisant la REalité II)

When you already hold a decision table, ELECTRE II does not sort alternatives into a core set the way ELECTRE I does. It is instead a ranking method that places every alternative in a single order. Its output is not a numerical score but a position: which alternative comes before which, which are tied, and which cannot be compared on any grounds.

Roy and Bertier proposed it in 1973 as a direct successor to ELECTRE I; its first application was a media-planning problem. It is one of the early members of the ELECTRE family and of the broader outranking tradition, and it is preferred for complex decisions with many stakeholders. It does not produce weights; weights are supplied from outside.

ELECTRE II inherits ELECTRE I's question of which alternative is not clearly beaten by any other. Establishing that one alternative outranks another still requires two conditions together: agreement across a sufficiently large share of the criteria, and no criterion where the disagreement (discordance, or veto) is severe. This is the non-compensatory logic inherited from ELECTRE I, in which a serious weakness on a single criterion can never be offset by strength elsewhere.

ELECTRE II's own philosophical contribution, however, lies in how it turns this logic into an order. The method does not make a single pass. It builds the same outranking relation twice, once with a "strict" concordance threshold and once with a "loose" one, giving a certain and a possible outranking relation. It then derives the order from two separate directions: once descending from the strongest alternative, once ascending from the weakest.

A firm preference or a tie is accepted only where the two directions reach the same conclusion. Where they disagree, the method does not hide it, but declares the pair "incomparable". This philosophical consequence makes ELECTRE II a kind of cross-validated ranking method. It aims to produce a full order, yet without blunting that aim it states openly where the data genuinely cannot discriminate.

The method proceeds through seven steps.

First, direction-sensitive concordance and discordance. For every pair of alternatives, the criteria on which one alternative is at least as good as the other are identified. Criterion direction matters here: on a "lower is better" criterion, the smaller value is superior. The weights of these criteria are summed to build a concordance value. Optionally, the ratio of the worst difference to the largest difference in the dataset gives a discordance value. ELECTRE II departs here from ELECTRE III and IV: it does not use graded preference and indifference thresholds (pseudo-criteria); the comparison is sharp.

Second, certain and possible relations. Two different concordance thresholds are set: a stricter one (s1) establishes a "certain outranking" relation, a looser one (s2) a "possible outranking" relation. An optional discordance ceiling can veto either relation: a sufficiently large disagreement on a single criterion invalidates the relation regardless of how high the concordance is.

Third, two graphs. A "strong" outranking graph is built from the certain relation, and a "weak" one from the possible relation.

Fourth, the descending order. Alternatives that mutually outrank one another in the strong graph (cycles) are grouped into a single class; the remaining classes are arranged from strongest to weakest. Alternatives within a class are further separated using the weak relation where needed.

Fifth, the ascending order. The same procedure is repeated symmetrically, this time building an order upward from the weakest alternative.

Sixth, the intersection of the two orders. The descending and ascending orders are compared. Where both agree, a firm preference or a tie is accepted; where they disagree, the pair is treated as incomparable.

Seventh, the result. The method returns the partial or full order built from these comparisons; it does not invent a single numerical score.

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

ELECTRE II's output is a position, not a score. You can say "A2 comes before A3", but not "A2 is such-and-such per cent better than A3"; a figure of that kind is not something the method produces. Three types of relation can arise: a firm preference (one alternative clearly ahead of another), a tie (two alternatives in the same rank), and incomparability (the case where the two directions disagree and the dataset cannot separate the two). Incomparability is not an error; it is part of the method's honesty mechanism. Changing the thresholds may or may not turn it into a firm preference.

Thus instead of writing:

"ELECTRE II placed A2 first, with the rest firmly second and third"

the report should read:

"With the chosen thresholds, A2 comes out ahead in both the descending and ascending rankings; the order between A3 and A1 is clear at these thresholds, but might become incomparable with a different pair of thresholds"

ELECTRE II works with crisp data: one number per cell. You need alternatives in rows, criteria in columns, direction information for every criterion, weights summing to 1, and two concordance thresholds (s1 > s2). An optional discordance (veto) ceiling can also be set.

It does not produce weights, it asks for them; they can be drawn from sources such as AHP, BWM, Entropy or CRITIC. DecisionMind holds five ELECTRE II family members alongside the base method (including fuzzy, intuitionistic fuzzy, multi-polar fuzzy and Pythagorean fuzzy extensions). It works with two to thirty alternatives, and is most comfortable with three to fifteen; it accepts two or more criteria, and three to twelve criteria are recommended.

ELECTRE II is a sound choice if you want a very poor performance on one criterion never to be offset by good performance on the others. It also serves when you need a full order rather than the core set that ELECTRE I produces. Its typical territory is complex decisions in which the interests of multiple stakeholders conflict.

There are situations in which it should not be used. If graded preference and indifference thresholds are needed, the pseudo-criterion logic that treats small differences as "unimportant", ELECTRE II's sharp comparison falls short. If a weakness on one criterion may reasonably be offset by other criteria, a simpler ranking method is enough. If a firm decision is required for every pair, that is, an incomparable outcome cannot be accepted, a method that always yields a full order is more suitable.

A full order plus veto logic (non-compensatory) is wanted → ELECTRE II

Not an order but a "set of front-runners" is enough → ELECTRE I

Graded preference/indifference thresholds (pseudo-criteria) are needed → ELECTRE III, ELECTRE IV

A weakness on one criterion may be closed by another → TOPSIS, SAW

A firm decision is required for every pair, an incomparable outcome is not wanted → PROMETHEE II

ELECTRE II's most distinct strength is that it keeps ELECTRE I's veto-like, non-compensatory logic while producing a usable full order instead of a core set. Second, it builds the order not in a single pass but from two independent directions (descending and ascending), and reaches a firm decision only where the two agree; this cross-validation is regarded as one of the most honest mechanisms in the outranking family (Govindan and Jepsen, 2016). Third, it does not conceal an incomparable outcome: where the data genuinely cannot separate two alternatives, it says so openly rather than hiding this behind an invented rank number. Fourth, as one of the early, foundational members of the ELECTRE family, it laid the groundwork for later members such as ELECTRE III and IV, whose threshold structure is more flexible (Roy, 1991).

Its limitations stem from the same structure. First, the choice of the two concordance thresholds (s1, s2) and of any discordance ceiling determines the result directly. Accepting these thresholds as "standard" values without regard to the data is a common risk and can change the result substantially. Second, unlike ELECTRE III/IV, it does not use graded pseudo-criterion thresholds. As a result, even differences as small as measurement noise, close to the threshold boundary, can be sharply split into "concordant/discordant".

Third, although the incomparable outcome is honest, it does not give a directly applicable order for some pairs; in practice this may require a second decision mechanism. Fourth, the ELECTRE family in general is criticised for the subjectivity of parameter selection and for difficulty scaling to large numbers of alternatives (Govindan and Jepsen, 2016).

The most common mistake is setting the two concordance thresholds (s1 and s2) equal to or very close to each other. This makes the distinction between "certain" and "possible" outranking meaningless and disables the method's dual-direction check. A second mistake is either never using the discordance (veto) ceiling, or setting it very strictly without regard to the data, and then being surprised that "the method could not separate any alternative". If the ceiling is too strict, no pair can enter an outranking relation and all alternatives end up in a single class.

A third mistake is confusing an incomparable outcome with a tie (indifference); the two carry different information. A fourth is treating only the descending order as final without intersecting it with the ascending order. A fifth is assuming that graded preference thresholds, as used in ELECTRE III/IV, also apply here; ELECTRE II's comparison is sharp.

The governing principle is this:

The order ELECTRE II produces is a consequence of the two chosen concordance thresholds and, where used, the discordance ceiling; the more sensitive the order is to these parameters, the more the decision should be reported as open to challenge.