An eigenvector power rating treats a football schedule as a strength matrix and ranks teams by that matrix’s dominant eigenvector — the same object, guaranteed to exist by the Perron-Frobenius theorem, that underlies Google’s PageRank.
From a schedule to a strength matrix
Building the matrix starts with a simple idea: estimate, for every pair of teams that played, a number representing how convincingly one beat the other. James Keener’s method, which applies the Perron-Frobenius theorem directly to ranking American football teams, defines a matrix entry A_ij as an estimate of the probability team i would beat team j, typically built from the ratio of points scored between them. One common form, using Laplace’s rule of succession to avoid dividing by zero when a matchup has few games, is
where S_ij is the number of points team i scored against team j across their meetings. Every off-diagonal entry of the matrix is filled in this way for every pair of teams that actually played; pairs that never met get a zero (or are left out of the graph entirely, depending on the implementation).
Why the dominant eigenvector is the rating
Once the strength matrix A is built, the rating vector r is defined as the solution to
which says r is an eigenvector of A, and lambda is its corresponding eigenvalue. Among all of a matrix’s eigenvectors, the one associated with the largest eigenvalue (the dominant eigenvector) is the one used as the rating, because of a specific guarantee: the Perron-Frobenius theorem states that for a non-negative matrix whose underlying graph is irreducible (a single connected component — every team reachable from every other team through a chain of games), the dominant eigenvalue is unique and its eigenvector can be chosen with strictly positive entries. That positivity matters practically — it means every team gets a sensible, comparable strength value, with no team landing on a zero or negative number that would be meaningless as a rating.
This is structurally the same guarantee that makes a continuous-time Markov chain rating’s equilibrium well-defined: both methods lean on Perron-Frobenius, applied to two different kinds of matrices built from the same underlying schedule graph — a rate matrix in the Markov chain case, a strength-ratio matrix here.
Computing it by hand with power iteration
Solving for an exact eigenvector normally calls for numerical linear-algebra software, but there is a simple iterative method that converges to the dominant eigenvector using nothing more than repeated matrix multiplication — well within reach of a spreadsheet’s array formulas, which is exactly why eigenvector power ratings are a common “build it in Excel” exercise. Power iteration works like this:
- Start with any non-negative guess vector, for example all 1s.
- Multiply the current vector by the strength matrix A.
- Divide the result by its largest entry (or its sum) to keep the numbers from growing or shrinking without bound.
- Repeat steps 2 and 3 until the vector stops changing meaningfully between iterations.
Each pass pushes the guess vector closer to the true dominant eigenvector, because repeated multiplication by A amplifies the component of the vector that lies along the dominant eigenvector’s direction relative to every other component. In a spreadsheet, this is a column of numbers, a matrix-multiplication formula referencing the strength matrix and the previous column, and a normalization step — copied across a few dozen columns until the values converge to a stable ranking.
A small worked example
Take three hypothetical teams — Kestrel, Marsh, and Talon — who played a short round robin, with Kestrel clearly the strongest. Suppose a simplified strength matrix, with entries representing the estimated probability the row team beats the column team:
| Kestrel | Marsh | Talon | |
|---|---|---|---|
| Kestrel | — | 0.70 | 0.80 |
| Marsh | 0.30 | — | 0.60 |
| Talon | 0.20 | 0.40 | — |
Starting the power-iteration guess vector at [1, 1, 1] and multiplying by this matrix (treating blank diagonal entries as 0) gives [1.50, 0.90, 0.60] after one pass; normalizing by the largest entry gives [1.00, 0.60, 0.40]. A second pass, multiplying [1.00, 0.60, 0.40] by the matrix again, gives [0.74, 0.54, 0.44], which normalizes to [1.00, 0.73, 0.59]. After a few more passes the vector settles close to [1.00, 0.68, 0.53] and stops moving meaningfully between iterations — Kestrel rated clearly highest, Marsh second, Talon last, matching the matrix’s built-in assumption that Kestrel dominates both opponents.
What the strength-matrix framing assumes, and where it can go wrong
An eigenvector power rating is only as good as the strength matrix feeding it, and that matrix embeds real assumptions worth naming. Using a score ratio (points scored between two teams) as a proxy for win probability treats a 35-7 result and a 21-3 result as roughly equivalent signals of dominance, even though the raw score gap is very different — the exact same “how much does margin of victory matter, and how” question that separates a Massey least-squares rating from a margin-blind method like the Colley matrix resurfaces here, one layer down, in how the matrix entries themselves get defined.
Connectivity is the other practical constraint. The Perron-Frobenius guarantee of a unique, all-positive dominant eigenvector depends on the underlying graph being a single connected component — every team linked to every other team through some chain of common opponents. A team (or a small group of teams) that only ever plays within its own isolated bubble breaks that guarantee, the identical failure mode that affects a continuous-time Markov chain rating’s equilibrium and for the identical mathematical reason: both methods lean on the same theorem, applied to different matrices built from the same schedule graph.
Where this sits among the other methods in this series
Keener’s eigenvector method is one more entry in the same family tree as Massey least squares and a continuous-time Markov chain rating — all three build some matrix from the schedule and extract a rating from it, but the extraction mechanism differs at a fundamental level: Massey minimizes squared regression error, the Markov chain approach finds a rate matrix’s null space, and Keener’s method finds a strength matrix’s dominant eigenvector. None of the three is a strict upgrade over the others; research comparing PageRank-style, Massey-style, and pure eigenvector methods for rating and ranking stability has found the results depend on the sport and dataset rather than any one method dominating universally.
Building any of these from real data starts the same way: a schedule of completed games. nflverse-data’s games.csv release provides that for the NFL back to 1999, from which a strength matrix like the one above can be assembled per matchup before running power iteration or a proper eigenvalue solver (numpy.linalg.eig handles this in a few lines once the matrix is built). The rating-consistency discussion picks up directly from here, comparing how closely an eigenvector rating’s ranking tracks the other methods in this series across a real season.
Frequently asked questions
What is an eigenvector power rating?
It is a team rating derived from the dominant eigenvector of a strength matrix built from game results. Each entry of the matrix estimates how strongly one team beat (or would beat) another, and the eigenvector that solves Ar = lambda r, guaranteed to exist and have non-negative entries by the Perron-Frobenius theorem, becomes the rating vector.
Who developed the eigenvector approach to ranking football teams?
James Keener wrote about applying the Perron-Frobenius theorem to ranking American football teams, building a rating method where the strength matrix entries estimate a win probability between two teams and the rating vector is the matrix's dominant eigenvector — the same linear-algebra object underlying Google's PageRank, applied to a schedule instead of a hyperlink graph.
Can an eigenvector power rating be computed in a spreadsheet?
Yes, using power iteration: start with any non-negative guess vector, repeatedly multiply it by the strength matrix, and renormalize after each multiplication. The vector converges toward the dominant eigenvector after enough iterations, and every step is just matrix-vector multiplication and division — well within what a spreadsheet's array formulas can do without specialized eigenvalue software.
What does the Perron-Frobenius theorem guarantee for a power rating?
For a non-negative, irreducible matrix (meaning the underlying graph of games is a single connected component), the theorem guarantees a unique largest eigenvalue with an eigenvector that has strictly positive entries. That positivity is what makes the eigenvector usable directly as a rating — no team ends up with a nonsensical negative or zero strength value.
How does an eigenvector power rating differ from a least-squares rating like Massey's?
Massey ratings come from a regression that minimizes squared prediction error on margin of victory. An eigenvector power rating instead treats the schedule as a strength matrix and finds that matrix's dominant eigenvector directly — no minimization, no error term, just solving Ar = lambda r. Research comparing the two families has found PageRank- and Massey-style methods can be more stable than a pure eigenvector approach in some settings, though results vary by dataset.
Does an eigenvector power rating need every team to have played every other team?
No, but it does need the schedule graph to be a single connected component — every team linked, directly or through a chain of common opponents, to every other team. Without that connectivity, the Perron-Frobenius guarantee of a unique positive eigenvector breaks down, the same requirement that governs a continuous-time Markov chain rating's equilibrium.