Wednesday, March 6 at 11:00 PM ET
Early last week, my brother and his wife were hanging with me and my wife. College basketball was on (at least on one of the TVs). From the wives' perspective, it was on "in the background." At some point near the end of a game or two, my sister-in-law looked up at a screen and asked why so many basketball games were close. My brother gave a relatively logical answer about only ever being able to score one, two or three points (or zero) in each possession and that each game allowed either team to gain a similar number of possessions. I gave my normal take on such questions, which is to say, "that is likely more perception than reality."
And then I went to Sloan Sports Analytics Conference at MIT this weekend. And then I watched basketball.
And then I dug deeper into our results. And then I started working on a major engine revision that could prove beneficial for all sports.
In my Bracketology class at the University of Cincinnati, I teach our students that basketball can be boiled down to shooting, turnovers, fouls and rebounding (what Ken Pomeroy has coined the "Four Factors"). Every possession ends in a foul, turnover or shot and all possessions can be extended by offensive rebounds. All ten players on the court, the officials and the coaches play a role in the general likelihood of each. This is essentially how simulations engines work. We determine who has the ball to end the possession, who is defending the player, whether he is likely to shoot, be fouled or turn the ball over and how the other eight players modify those chances. If a shot is missed, we figure out who gets the rebound. There are situational factors that we layer in that have to do with coaching and end of half and end of game scenarios, but essentially, that's basketball.
I glossed over that last part. I should not have.
(We don't usually cover performance in this space, but this has a very important point. All numbers below come from our transparent ResultsFinder Database for the 2012-13 college basketball season. Each of the subsets where the percentages only are given includes at least 400 games from this year.)
This season in college basketball, our playable (52.4%+ confidence) "Upset Watch" picks are 118-90 (57% ATS). Our normal or better (greater than 57% confidence) "Upset Watch" plays are 59-40 (60% ATS). In games where we like a 0 to 2.5 point favorite to cover, we are 69-52 (56% ATS) on playable picks and 20-10 (67% ATS) on normal+ picks. When taking 3 to 9.5 point favorites to cover, we are 49% accurate ATS on playable picks and 46% ATS on normal+ plays. Also, when picking the favorite to cover (at all, not just when playable), normal+ O/U plays are 49% accurate. And, when taking the underdog to cover, normal+ O/U are 61% accurate.
We have long been promoting the general success of our "Upset Watch" plays and normal+ over/unders in the ResultsFinder highlights and in emails. As it turns out, this "trend" can be seen with all of our sports as well. We are consistently profitable (very much so in college football and MLB totals - and in the postseasons of all sports due to motivation) with strong O/U plays and upset picks across the board.
Sitting in on the presentation by Yahoo! Research Scientist, Dr. Justin Rao of a research paper at SSAC 2013 - and subsequently watching the Miami @ Duke, Virginia @ Boston College and Michigan State @ Michigan games, led me to believe why this is the case. The paper that was presented (Side note: If going to SSAC, go see as many research papers as possible - this is where the real education occurs.) was entitled, Live by the Three, Die by the Three? The Price of Risk in the NBA. The paper discussed how teams in the NBA address game situations depending on time remaining and point differential. Not surprisingly, teams losing substantially take more threes and teams winning by low margins do not take many threes.
Three point baskets have greater risk than two point baskets because a team trades in a possession for a shot and the probability of a successful possession from a three point shot is lower than the probability of a successful possession on a two point shot (even if the expected value of points from a three point shot is greater). Teams nursing leads do not like that risk. Teams that are trying to battle back from deficits are willing to assume that risk... I have not said anything groundbreaking yet. It's not the three point strategy that provided the "ah-ha" moment. It's this...
From the empirical data, Rao and his co-author Dr. Matthew Goldman, were able to illustrate that an average NBA team is 20% more efficient when losing by ten points than when winning by ten points. Holy #*$%! NBA fans may have thought something like this, yet now we have play-by-play data to back it up and then some. To put that into perspective, the difference between the most efficient NBA offense (Oklahoma City) and the least efficient (Charlotte) only crosses about a 15% variance. In its most basic sense, a completely average NBA team plays as well as Oklahoma City when losing by ten points and like Charlotte when winning by ten points. Wow.
So what are the takeaways?
The most significant to me (and likely you) is that we now have actual data and well-constructed research to directly employ into our simulation engines (ASAP). As alluded to, simulation engines have many situational pieces that look at the importance of shooting threes vs. twos in certain situations. We also have directly factored in performance by quarter for each teams, which goes hand-in-hand with expected playing time for players by quarter (the best and the worst teams are usually the teams that play best in the fourth quarter as really good teams play really well usually and put their opponents away when necessary and really bad teams become more efficient when trying to dig out of large deficits late in games), but this takes that to the next level. Obviously, this has very important ramifications for Live ScoreCaster (now totally free), yet it also helps with the performance we just discussed because...
While it may be true that Michigan is more than 4.5 points better than Michigan State on Michigan's home floor, that does not necessarily mean that the Wolverines are very likely to win by five or more points - due to game factors related to how teams play in certain deficit situations. The simulations have proven great and assessing the overall pace of most games with O/U performance, when the wrong team is favored, or when there is a clear favorite in what is expected to be a very close game. This should not impact those considerations. Instead, difficulty has arisen in games with short-to-medium lines - where there is one, obvious, agreed-upon favorite, yet the underdog has a decent chance to win (especially in around 4.5 to 9.5 point spread games in basketball and around 7.5 to 12.5 in football).
And that, Tiffany (my sister-in-law), is (a reason) why there are so many close games. Parity exists partially because there are so many evenly matched teams and partially because there are so few elite or terrible teams that result in the blowout games to which this analysis does not apply.
Also, there is one other, very important takeaway from this analysis that is not as obvious. If teams that need to be risky are 20% more efficient than teams that minimize risk, taking more (calculated) risks leads to improved expected performance. This is something that poker players, Madden video game players and a few coaches and front offices figured out many years ago. Know the odds and exploit them to go big.
As usual, if you have any of your own comments about this article or suggestions about how to improve the site, please do not hesitate to contact us at any time. We respond to every support contact as quickly as we can (usually within a few hours) and are very amenable to suggestions. I firmly believe that open communication with our customers and user feedback is the best way for us to grow and provide the types of products that will maximize the experience for all. Thank you in advance for your suggestions, comments and questions.