One thing that has always bugged me about Bradley Terry models is that it doesn’t model rock-paper-scissors. To recap the Bradley Terry model posits that each entity has a strength \(\gamma_{i}\) and when competing the probability of victory for entity a over entity b is \[exp(\gamma_{a}) / exp(\gamma_{a}) + exp(\gamma_{b})\] or to make it clearer logit(γa - γb). ...