I had a website about this for several years called beatpaths.com . Some of the graphs for it are still up, and another person took the idea and continued auto

I had a website about this for several years called beatpaths.com . Some of the graphs for it are still up, and another person took the idea and continued auto generating similar graphs over at beatgraphs.com. It was fun, but not really revolutionary. You'd often end up with cycles, and I had a system to discard the "beat loops" (smallest loops first) until you'd end up with a DAG. You could make predictions by saying Team A would probably beat Team B if it had a path to team B in the graph (the basic transitive property), but still, the performance wasn't better than Vegas. I went as far as coming up with some custom tsort algorithms; various tiebreakers, to come up with a full power ranking that fully respected the implied orderings in the graph. In pro football, there's a dummy algorithm you can use to test your homemade custom snazzy algorithm. The dummy algorithm is: pick the team with the better record, and if they tie, pick the home team. If your algorithm doesn't match it, then your algorithm isn't worth a whole lot. My tsort algorithms didn't regularly beat it. If I had altered the dummy algorithm to just say: 1. Pick the team with the better record, 2. pick the team that has a beatpath to the other, 3. pick the home team - then I probably would have beat the dummy algorithm by a very little bit. But still, not close to Vegas performance. In college football, trumpeting 88% is dumb because college football regularly has crap matchups. I had some full beatgraphs to college football, too - they were huge and loosely connected, and its performance wasn't surprisingly good or anything.