I am posting my bowl season results with only the championship game remaining tonight. Here is how the model performed by game.
|54%||Northern Illinois||40||17||Fresno State||46%|
|45%||Navy||14||35||San Diego State||55%|
|59%||Air Force||14||7||Georgia Tech||41%|
|55%||West Virginia||7||23||NC State||45%|
|57%||Notre Dame||33||17||Miami (FL)||43%|
|51%||South Carolina||17||26||Florida State||49%|
|39%||Mid Tennessee||21||35||Miami (OH)||61%|
The expected model accuracy was 58%. Actual model performance was 59%. If we look at the distribution of expected results, we get the following. Blue bars represent relative probabilities of the model predicted the correct results for the number of games on the x-axis. 34 games have been played. The dark line represents the expected probability from guessing each game (50%). The model is clearly a little better than guessing.
The model was not completely uniform. It predicted better at the higher probabilities and not as well near 50%.
If we consider the results as the percentage of points score by the predicited winner, then the model correlates well to the result.
So, going into the championship game, I don’t think I could do any better than a coin toss on picking the winner. I would expect the game to go right down to the last couple of minutes of the 4th quarter.
Still: GO DUCKS!