Scoring Rules as Least-Squares Estimators

Abstract

Kawada (2018) proved that a scoring rule represented by a strictly positive score vector is equivalent to the corresponding cosine similarity rule. The original proof relies on a direct analysis of the cosine similarity optimization problem. In this note, we present an alternative, simpler proof based on a basic least-squares characterization. For any fixed preference profile and score vector, the arithmetic mean of the individual score vectors uniquely minimizes the sum of squared Euclidean distances to those vectors. It then shows that, for any strictly positive affine representative of the given score vector, the optimal solutions of the cosine similarity problem are precisely the positive multiples of the arithmetic mean of the corresponding score vectors. Since the two arithmetic means are positively affinely related, they induce exactly the same social ranking. This result provides a transparent geometric interpretation of scoring rules and clarifies why the cosine similarity rule necessarily coincides with the corresponding scoring rule.