What is this?

This page displays predictions for games in the 2019 CFB season. These predictions come from a model built on historical college football play by play and game data in order to simulate upcoming games. All data is from collegefootballdata.com.

1 Week 15 Predictions

This table displays the results of simulating week 15 of the 2019 CFB season 1000 times. These simulations are run one week in advance, using information available in the lead up to week 15. For the first week of the season, team ratings are based solely on information from previous seasons.

Quality is the (harmonic) mean of the two team’s overall rating, with 100 indicating a game between highly rated opponents. Pr(HomeWin) is the percentage of times that the home team won the game across every simulation.

2 Results Through Week 14

How has the model performed so far in 2019?

2.1 Assessment

The model predicts the probability that the home team will win and their expected margin of victory (or defeat). I assess the model’s performance by assessing the accuracy of its predictions (win/loss), the logloss of its probabilities, and the mean absolute error of the margin.

			Outcome	Margin
2019	1	85	0.91	0.31	11.8
2019	2	74	0.84	0.35	13.9
2019	3	68	0.82	0.41	15.2
2019	4	58	0.76	0.47	12.1
2019	5	55	0.84	0.40	13.7
2019	6	48	0.69	0.58	12.2
2019	7	53	0.68	0.60	13.1
2019	8	62	0.71	0.54	13.2
2019	9	55	0.69	0.59	14.4
2019	10	48	0.75	0.47	14.1
2019	11	48	0.60	0.58	13.6
2019	12	55	0.84	0.45	11.5
2019	13	64	0.78	0.44	12.0
2019	14	64	0.75	0.49	12.5
Season Average:	0.77	0.47	13.1

2.2 Results

The following table shows the results of games played in previous weeks next to the model’s predictions.

3 Predicting the Rest of the Season

This table displays the results of simulating the rest of the 2019 season 1000 times. These simulations are run weeks in advance, where the remainder of the season is simulated using only the information as of week 14.

For example, we simulate the entire season before any games are played by simulating week two based on the (simulated) results of week one, then simulating week three based on the (simulated) results from week two. As a result, predictions for future weeks tend to be less accurate than predictions made one week in advance and are likely to change as we get closer to the actual game.