Quote:
Originally Posted by tarheels2222
Yes, I believe so, as that's really its only baseline. So there is certainly some nuance, especially in a single sample. But over the long run, my understanding is that the data should all converge to create the estimated WP delta.
A few posts above, I posted a link with some more info on it. There is also an Athletic article link in that info that gives more context.
Below are a few of the problems using league averages.
1. The data is old. The game changes every season, with new players, new coaches, new rules, new strategies, new tactics, etc.
2. The game is not played with "average" defenses and offenses. The analytics folks are lazy and use averages because it generates the largest set of data, thus reinforcing one's belief that it is the most statistically accurate projection. Would using league averages be appropriate if the QB cannot run the ball in, thus providing the defense with a huge edge? Would it be correct to apply league averages to arguably the top defense ITL?
3. As alluded to earlier, historical games have been played sub-optimally. Think of a chess match with 2 world class players. Would a top player play the move that beats "average" opponents? Or would the player need to think more about deeper lines to counter the ability of the opponent to respond to "average" moves?
4. Of course there is convergence to the mean with large data sets. But why would that be of any interest in a specific situation? Averaging out all future results will converge to a mean, but the mean is not known, and there are plenty of valid reasons why averages would not apply to a specific situation. In reality the data set offers a multitude of means, each appropriate for the specific situation. Using a mean of means would be analogous to a model projecting rate of traffic deaths using collision speed and ignoring make/model of the vehicles.
5. Subsets using more intelligent historical data to reduce the averaging effect could improve accuracy of the projection, but at a cost of using smaller data sets.
6. Also noted earlier, using averages almost always results in the "chicken or the egg" scenario. Without seemingly "bad" decisions mixed in with "good" decisions, there would be no average to calculate because all decisions would be the "average" decision.
7. Using averages to determine the proper play goes against game theory. For example, a team should not always punt from the 50 because that is giving away too much to the opponent. Without the threat of an alternate play, defenses can play more efficiently. Another example: runs on 2 pt conversions succeed at a higher rate than passes. But does that mean that all teams should always run the ball because the average run is more effective than the average pass?
I am quite familiar with constructing models for game projections. I can tell you that using a simple model to project win probability based solely on historical averages of point margin and time remaining is weak, and should not be part of the conversation beyond a rough first order approximation.
A drive or play model that takes into account compositions of teams, playing surface, weather, injury rates, correlations of previous plays to future plays, coaching tendencies, etc. that can be iterated over millions of times will produce a much better result than using generic league averages compiled over years of historical data.
Quote:
Originally Posted by Zimmer4141
I am kind of shocked by the WP that the model spits out for the end of half decision.
4th and 2 is essentially a 2PC attempt which are converted 48.2% of the time. So even by an Expected Points scored calculation, going for it is 3.37 expected points where a FG is basically 3.
I wouldn't be surprised if end of half scenarios are really hard to program into the model, and it's calculating it with the added value of SF getting the ball at their 2 when that doesn't hurt them at all.
The 1H play was on the 3, and the historic average for TD is 33%. Assuming the FG is 98%, the break even going for TD over FG using historical averages is 42%.
0.98 x 3 / 6.94 = 0.424
where 6.94 is the expected points from a TD and 94% XP. This could be improved slightly by going for 2 pts assuming the offense has an effective 2 pt play.
Of course this ignores win probability and falls into the averaging trap. If Lions indeed had a 4th and 3 play that could be expected to succeed higher than 42%, and their models showed that points were not at a premium (thus improving WP), than going for TD there would be the correct play.
As the game continues, kicking FG becomes the better play assuming same point margin.
Quote:
Originally Posted by Onlydo2days
I think in the mahomes 1st SB year, houston was up 21-0 and kicked a FG on 4th and 1 at the 20 and then the chiefs just killed them after that and people said they needed to know their D sucks, stay aggressive and go for it.
4th and 1 is different than 4th and 2/3 but taking 3 up 21-0 makes more sense than taking 3 up 24-10 on the road against an offensive juggernaut when your pass D sucks
Texans had the ball 4th and 1 from the 16 up 21-0 mid 2Q. There was a long thread here with stinkypete being the sole defender of the FG decision. According to a suzzer post Texans were over 90% to get 1st down, actually higher than a successful FG.