Searching For Wins
Searching For Wins
Bowl eligibility now looks unlikely. How unlikely? I built a new multiple regression model to find out.
Note: Data is sourced from CFDB and ESPN, except where noted. All PFF grades are preliminary, and current as of 5:00 a.m. Sunday morning (October 2). These grades can, and likely will, shift a bit, but the broader takeaways should hold true.
For the last few weeks I have been working on a multiple regression model to predict the scores of Virginia Tech games. For each of the first five games, I included a “Gut Feelings” section in my preview article. The goal of that section is add qualitative data to the all the quantitative metrics and present a balanced estimation of how I expected the game to proceed. The problem was that I did not have a baseline estimation of the score. The model changes that.
Model Details
The dataset that powers the model includes all Virginia Tech games for which there are data, starting with the 2003 season. Games for which there are insufficient data (i.e., against FCS opponents) are not included. The multiple regression model to predict VT’s point total includes the following variables:
VT pre-game ELO
Opponent pre-game ELO
Healthy VT quarterback
VT home/away
VT offensive and opponent defensive explosiveness
VT offensive and opponent defensive power success
VT offensive and opponent defensive PPA
VT offensive and opponent defensive stuff rate
VT offensive and opponent defensive success rate
VT offensive and opponent defensive field position predicted points
The same variables, with the advanced stats in reverse (opponent offense vs. VT defense), are included in the regression to predict the opponent’s point total.
Accuracy
Including the North Carolina game, the model accurately predicted the winner in 177 out of 237 FBS games, for a total accuracy of 74.7%. The model had two down years at the end of the Beamer era, 2013 and 2014, in which it went 6-6 and 5-7. Excluding those two seasons, the model has gone 164-45 (78.5%) since 2003. In comparison, Nick Saban is 185-25 (88.1%) at Alabama. So, the model is not quite Nick Saban, but it is still pretty darn accurate, especially considering that the goal is to predict the point totals of the two teams. Predicting the winner is a secondary goal. (Incidentally, I built a separate multiple regression model with all the variables included for both teams’ point prediction to simply predict whether or not VT will win. That model has demonstrated a 79.4% accuracy rate, so an increase of 5 percentage points over the point predict model. This model is better at predicting outcome, though, than point totals - I checked.) Here is how the point predict model did through the years in predicting the winner:
Again, though, the real value in this model is predicting scores. Let’s do a quick survey of big games with somewhat unexpected (at least to the fan base) results as a sort of smell test:
2005 ACC Championship vs. FSU - Predicted: win 23-22, Actual: loss 22-27
2007 at LSU - Predicted: loss 20-25, Actual: loss 7-48
2007 Orange Bowl vs. Kansas - Predicted: loss 15-22, Actual: loss 21-24
2010 vs. Boise St. - Predicted: win 32-29, Actual: loss 30-33
2011 Sugar Bowl vs. Michigan - Predicted: loss 26-28, Actual: loss 20-23
The model went 3-2 in predicting the outcome of those games, but take a look at the score predictions. It was very close in four out of five games, not just in point differential, but also in point total.
Virginia Tech entered the 2005 ACC Championship 10-1 and ranked #5 in the country. FSU was 7-4 and ranked #22. The model was way more accurate than just about any human was going into that game (VT was a 14-point favorite).
The model picked Virginia Tech to lose to LSU in 2007 in a matchup of top-10 teams, but was way off on the margin of victory. I’m not sure anyone saw that disaster coming. But the model rebounded with a very good call on the Orange Bowl defeat to Kansas. Kansas! The model was also very close on both the Boise St. and Michigan scores in 2010 and 2011, respectively. Boise St. scored the go-ahead touchdown in the last minute of the game, and Michigan won in overtime.
Fast Forward to Today
The model predicted North Carolina would defeat Virginia Tech 38-12. In my pregame update, I had predicted UNC would win 38-17. (Note: I completed the model after the game, so it did not influence my pre-game expectations). In actuality, the Tar Heels won 41-10. So, the model was more accurate in its prediction than I was.
Virginia Tech is currently 2-3 overall and 1-1 in the ACC, with three more tough games in October, then four easier ones in November. Bowl eligibility is still possible, but increasingly less likely. Based on current data (the model gets data refreshes each week, so these predictions are subject to change), the model foresees VT winning only one more game this season, 27-15 over Georgia Tech. Virginia Tech needs four wins to gain bowl eligibility. The three other games VT is most likely to win, according to the model, are:
at Liberty, predicted loss 26-28
vs Virginia, predicted loss 24-28
at Miami, predicted loss 24-31
That’s a tall order, but not impossible. The model does not expect the other games to be close:
at Pitt, predicted loss 17-49
vs Duke, predicted loss 19-39
at NC State, predicted loss 28-40
How to Win 4 More Games
The talent gap will not get fixed this year, but there are some things the Hokies can do to improve their chances of becoming bowl eligible:
Fix Parker Clements or bench him - in 360 snaps this season, Clements has graded out at 55.9 on pass blocking and 44.2 on run blocking. Those grades are not conducive to starting at right tackle in the ACC or really on any FBS team. Last year, his grades were 63.8 and 75.0, respectively. While Clements hovered around 50 in both categories this past weekend, every other lineman earned a grade of at least 60 in both pass and run blocking against North Carolina.
Get healthy for November - there is a chance Virginia Tech could pull the upset at home against Miami, but VT will face steep odds at Pitt and NC State. There is no sense playing dinged up guys in those games. Let them heal up for the stretch run.
Figure out the playing rotation at TE and WR, then stick with it - one week it’s Connor Blumrick, the next Christian Ross, then here comes Dae’quan Wright out of nowhere. The unpredictable rotation at these positions seems to be contributing to the disjointed passing game.
Get more pressure on the quarterback - the pass rush has really dropped off the last two weeks:
Virginia Tech, as a team, has the seventh highest pass rushing grade in the nation, per PFF. Still, as the level of competition has increased, the pressure applied on opposing quarterbacks has declined. In the first three games, Tech was pressuring the quarterback on every other dropback. That pressure rate has been halved in the last two games, with the Hokies defense only pressuring the quarterback on about 25% of dropbacks. VT should, and I expect they will, shift snaps from the DE combo of Garbutt/Griffin to Nelson/McCray.
When healthy, Garbutt is the best DE on the team, but I am not convinced he is healthy. In the last two weeks he’s notched just 1 sack and 3 hurries while looking slow out of his stance.
The Takeaway
If it feels like the season is on the verge of getting away from the Hokies, (the numbers would) that’s because it is. More losses are coming, and some of them may be blowouts. Still, there are winnable games remaining on the schedule. If the Hokies could steal a game in October, I would feel a lot better (but still not good) about their chances of making it to a bowl this season. However, if they enter November with two wins and on a five-game losing streak, would you, or anyone else, bet on them running the table?