Key Tenets

Relevance-based prediction rests on three key tenets

Relevance

• Relevance is a mathematically precise measure of the importance of a player in the training sample to the prediction of an outcome for a player of interest.

• Relevance is composed of two components, similarity and informativeness, both measured as Mahalanobis distances.

• Players from the training sample who are statistically similar to a player of interest but different from average are more relevant than those who are not.

• We form a prediction of an outcome for a player of interest as a relevance-weighted average of prior outcomes of players from the training sample.

Fit

• Fit measures the alignment of relevance and outcomes across all pairs of players that go into a prediction task.

• A given pair of players is aligned if they have similarly high relevance and similiarly high outcomes.

• Fit reveals the unique reliability of a specific prediction task.

• Fit also determines the uniquely optimal combination of predictive variables and observations for each individual prediction task.

Codependence

• Conventional prediction models assume that a chosen set of predictive variables is equally effective across all samples of players from the training sample, which is false.

• Conventional prediction models also assume that a chosen sample of players from the training sample is equally userful for all combinations of predictive variables. This too is false.

• By maximizing fit as a joint function of the predictive variables and players from the training sample, we determine their uniquely optimal combination for each prediction task.