Tuesday, January 31, 2017

Value Add Version 3.0 Explanation

Value Add 3.0 Introduction – December 30, 2015

Those wanting an easier explanation of the breakthrough Value Add Basketball 3.0 should click here. This post is a more detailed explanation of the new system and how it works. This Value Add Sports post gives background on from the 2011 invention of Value Add Basketball 1.0, and many of the articles written about the system.

The actual ratings of current players appear at www.valueaddbasketball.com, while team ratings are here and ratings for players going back to the 2002-03 season are here.

The programmers were about to shoot me before 2015 ended, but luckily more than 1,000 lines of code have finally been tested and calibrated to produce the first true run of Value Add 3.0.
VALUE ADD 3.0. The third version of Value Add Basketball calibrates each player’s stats to measure how much a player improves his team’s Offensive and Defensive Efficiency. A team of “replacement” players would be expected to score about 80 points per 100 trips down the court against an average defense. The 9.99 Offensive Add for LSU’s Ben Simmons indicates that even if all his teammates were replacement-level players, that he would improve the Offensive Efficiency Rating for the team from 80 to 90. His Value Add rating indicates that he would lower an average opponent from 102 (this year’s average) to 98.

The most lopsided team match-up would be University of Virginia’s team of players against Delaware State’s. To the starting point of 80, you add the total of Virginia’s players (36.22) to the Defensive Rating of Delaware State’s Defensive Ratings (11.22) to estimate Virginia’s Offensive Efficiency for that game would be 127 per 100 trips. So if both teams had 70 possessions during the game we would roughly estimate that Virginia should score 89 points.

For Delaware State, add the 80 replacement points to Delaware State’s 4.34 Offensive Rating and with Virginia’s NEGATIVE 10.22 Defensive Rating Delaware State would be estimated to score 74 points in 100 trips, or 52 points. So Virginia would be expected to win 89-52 on a neutral court. You add 2 points to the home team and subtract 2 from the road team, so if Virginia was at home it calculates a 91-50 Virginia win, at Delaware State an 87-54 win.

While Sagarin or Pomeroy will yield slightly more accurate predicted scores – since the sum of the players is not always equal to the overall value of any team – the advantage of this new calibrated Version is that it lets you to know the impact of an injured player or a player returning from injury.

NO SIMMONS COULD HAVE TURNED 11-POINT WIN TO LOSS. As shown above, if Simmons was injured we would expect LSU’s Offensive Efficiency to drop from the current 105.9 points per 100 trips to 95.9 per 100 after taking out Simmons 9.99. We would expect LSU’s defense to give up 103.7 points per 100 trips instead of their current 99.6. LSU took 84 trips down the court in the 119-108 win over North Florida. Simmons combined impact of 14 points per 100 offensive and defensive trips indicates he is worth 12 points to LSU in a game like that with 84 trips down the court – and LSU likely sees the 11-point win turn into a 1-point loss. (in fact, that night he was likely worth a little more as he scored 43 points – but most of the points a player scores would have been made up by other players)

PLAYERS GET 0.0 VALUE ADD WHEN OUT. By the same token, UNC became a much better team when Marcus Paige returned after missing the first six games. If you look at his offensive rating right now it is 4.70 per game. However, it is important to look under the “Year” column to see that he has only played in 7 of the team’s 13 games. Since he had a 0.0 Value Add in the six games UNC played while he was still injured, quick math tells us he has actually been worth an average of 8.73 offensive ratings points in his seven games. At the risk of making Northern Iowa fans angry, that rating in the 66 trips up the court against Northern Iowa adds an estimated 5.76 points to UNC’s score that night (8.73 per 100 trips times 66/100 trips that game) and gives UNC a 73-71 win instead of the actual 67-71 upset by Northern Iowa.

Several adjustments were part of the new system – some in response to specific criticisms after the mainly warm reception by NBA teams and sports media to the first system and others due to ongoing research that identified inaccuracies.  

STEALS WORTH 10.5% LESS. The initial Value Add equations assumed a player who caused steals also created additional turnovers by jumping in lanes. While steals are still a huge factor and underestimated by most fans, analysis over the years indicated we overestimated this impact and that in fact the number of “non-steal turnovers” a team forces does not appear impacted by the turnovers that ARE the result of steals (e.g. a team that plays sound defense without going for steals likely forces as many travels, shot clock violations etc. as a team jumping in the lane). As a result, Value Add 3.0 reduces the impact of a steal to 89.5% of the value under the initial system. In effect, all five players on the court at any given moment share most of the credit for each non-steal turnover whether than most of the credit going to players with more steals.

DEFENSIVE STOPS. We were able to more accurately measure the “other” defensive stops after removing the impact of defensive rebounds, blocked shots and steals and in quality control the overall defensive Value Add figures are much closer to each team’s defensive ratings.

DEFENSIVE MINUTES ALLOCATED (DOUG MCDERMOTT RULE). In allocating defensive stops, a factor is now built in which assumes the coach of a weaker defensive team is typically playing better defenders more – all other things being equal. Under the old formula Doug McDermott is the only player to have an OFFENSIVE RATING of over 7.0 during three seasons (JJ Redick is the only other player to do it even twice). However, because McDermott played at least 79% of the minutes every year and the Creighton defenses were poor the original system assumed he was giving up a big percentage of the buckets scored, but Version 3 lessens the impact of playing more minutes for a bad team to give the bigger penalty to players who get in the game less often on a bad defensive team.

GAMES PLAYED. Denzel Valentine’s 12-game Value Add is 13.49, but his value in the 13th game against Oakland – the first game missed with his injury, is 0.00 so his overall average falls to 12.45 to drop him from 4th to 11th place. The reverse is true in the Marcus Paige example above – a returning player is worth more per game once back from injury.

POSITION ADJUSTMENT. Version 3.0 sets a value so that the average D1 guard will always equal a 2.0 Value Add and the average front line player will also equal 2.0. When we update past years based on the new system the totals to guards raw scores will get a large boost because hand checking made it so much harder for guards. However, this season freedom of movement rules have helped guards get much better stats and thus they need no adjustment (SG*1), whereas the front line players actually needed to be increased by 1.5% (PF*1.015).

HIGH-, MID- AND LOW-MAJORS. High Major players do not play the full game in early season matchups against Low Major teams – so during conference play starts up the high major players gain about 2.0. To avoid the early season distortion all high major players receive an extra point (High+1) while Low Major players lose a point (Low-1) and Mid-Major players stay the same. This adjustment is phased out over the next two months.


TEAM VALUE ADD. Value Add is primarily a rating of individual players, but Version 3.0 yields individual Value Adds to tally more closely to the Efficiency Ratings of entire teams. If you go through the top Value Add teams all are within a few spots of where they would rank at www.kenpom.com until you get to Princeton – who for some reason is 40th based on their Value Add players but not even in the top 100 at www.kenpom.com. For some reason Harvard is the next team (88th in Value Add) that is much lower at www.kenpom.com, so maybe those smart Ivy Leaguers can figure out why Value Add seems to like their players.

Value Add Explanation of Version 1.0, with links to how the Offensive and Defensive Components are set up

(Jan 22, 2013 Summary of Value Add Version 1.0, with post notes)


For an explanation of the Value Add basketball rankings, you can always click on this post from the bottom of the www.valueaddbasketball.com database.  I must thank Sports IllustratedESPNNBC Sports and all of the other outlets who have covered Value Add.  A variation of this formula measures how likely a player is to be successful in the NBA, but the basic explanation of how and why it works follows.

Anthony Davis added 7.29% to Kentucky’s scoring with his offense, and took away -5.06% from opponent’s scoring, so his total impact on the score was 12.35%, the highest Value Add in the country in 2012. This means that if Kentucky would have lost a game 69-70 with a typical fourth or fifth man off the bench playing instead of Davis, then with Davis they win 74-66.
When you first pull up the database at www.valueaddbasketball.com, you will see the ratings of all of the basketball players for this season.  You can search by team, conference or a player’s name to see where specific players rank, or just scroll through more than 3000 players for each season.
You can also click on the other tabs to see players from the 2002 season through the projections for how good players should be in 2014, including new recruits.  This is probably enough explanation for most readers, but if you want more details, the ratings consist of three components:
1. The Offensive component is explained in this post, as we can measure with great precision how many points a player ad to his team’s score.
2. The Defensive component is not quite as precise, but measures a player’s ability to block shots, steal the ball, grab defensive rebounds and be part of a defense that prevents opponents from scoring in other ways.  In both cases, a players rating is measured against exactly how good each of an opponents’ offense and defense is, so the same player should have basically the same rating whether he plays for the national champion or the worst Division 1 team.
3. The Point Guard/Perimeter Defensive Rating (PG/Per) redistributes a small percent of the credit from post players who do not turn the ball over as much because they do not have to dribble as often and who grab more defensive rebounds because they do not have to play defense on the perimeter.  After extensive study, this figure was determined the most accurate way to fairly adjust ratings based on position, as explained in this post.
NBA Indicators Developed by Rob Lowe
Much like Ken Pomeroy at www.kenpom.com, Rob Lowe (who is much smarter than me) developed methods for measuring tempo free statistics as calculated in Dean Oliver’s book, “Basketball on Paper.”  While Oliver, and in turn Pomeroy, determined how to measure a dozen things a player does to help his team win, www.valueaddbasketball.com instead breaks down all ways a player helps his team into on figure that determines his overall value.
Lowe built on Value Add and other sources to develop calculations that measure how likely a player is to be successful in the NBA, and officials from several NBA teams met with me prior to the 2012 draft to review these valuations.  These evaluations are only available to select NBA teams, but we plan to provide real-time coverage of the NBA Draft (see last’s years first round and second round) as picks are revealed at Breitbart Sports.
Others have built on the Value Add system since it was developed.  Sports Illustrated used it to go back further in history and conclude that Duke’s JJ Reddick was the best Value Add guard since modern stats were introduced, and that research was picked up by ESPN. Basketball Prospectus then started to add to the equation with some additional calculations.
Big Apple Buckets out of New York was instrumental in the further development of a variation to calculate the value of players in the Ivy League and other low major conferences, and we have received follow-up questions from Ivy League schools and officials.  These ratings can be found by clicking on the “Low-Major” link in the bottom right corner.
In addition, these rankings are written about regularly on sites for teams like KentuckyArizonaNC StateBaylor and League sites for the Big Ten, Summit, Horizon and Patriot Leagues as well as the original www.crackedsidewalks.com page that focuses on Marquette and the Big East/Catholic 7.
One credibility test for the system occurred after an engineer developed the program to run the calculations, and a Marquette player, Jae Crowder, showed up as an All-American.  Statisticians never like to have a favorite when running numbers, and the fact that I covered Marquette and Crowder’s Value Add was so high even though he was not even an All-Conference player nor listed in the top 100 NBA prospects made me nervous.  However, several NBA team officials met with me and were very complementary that my calculations were “finding” several of the same sleeper prospects they had found through intense scouting, including Crowder.
I did later meet Crowder after his last game at Marquette, and was put at ease months later when the AP sports writers named him one of the top 10 players as a second team All-American and then the Dallas Mavericks drafted him and he was named named an All-Star in Summer League and made the Mavs.  While the numbers were backed up, I will give Crowder and the other players who backed up their Value Add ratings with great runs on the court assists for the success of Value Add.

I have also taken questions on Value Add on the radio, and spoken on the the statistics behind politics and sports.

Monday, January 30, 2017

Most Overvalued and Undervalued Stats

Here are the most overvalued and undervalued stats:

Undervalued

Missed Shots. It always amazes me when a fan talks about how clutch a player was to just keep shooting after starting a game one of eight because "he is their scorer" and hits a shot late and wasn't scared to keep shooting. If a player hits a game-winning shot after going one for eight, then the team should not have needed a last second shot. In many cases a trigger happy player costs his team the game.

Live ball Turnovers. Unfortunately we do not have a breakdown of how many live ball turnovers (steals) a player gives up, usually resulting in the opponent having a chance for a fast break. An out-of-control player who gives the other team the ball on a break several times en route to scoring some points is often hurting his team.

Defensive Points Allowed. I introduced this concept back in my historic Marquette book, but fans don't seem to grasp that the points a player scores is relative to how many points his team needs to score to win the game. Pomeroy lists %Pts to show what percent of his team's scoring a player accounted for his each game, but if he would just flip that to divide by points allowed. If a player scores 15 points and the opposing team his held to 50 points, then he has scored 30% of the points his team needs to win or at least force overtime. If a player scores 20 points and his team gives up 100 points, he has only scored 20% of what his team needs to win. This simple state (Points / Opponents Team Points) would fix so many distortions by taking into account tempo as well as defense..


Overvalued

Points per game. People seem to ignore missed shots, turnovers and poor defense if you have a 20 point game.

Free Throw percentage. Commentators go on and on about a team with a 71 percent free thrown percent not wanting to get into a free throw shooting contest with a team shooting 78 percent from the line. OK, so it can be a tie-breaker. If every other aspect of the game is played completely even and both teams shoot 14 free throws, the better shooting free throw shooting team will project to win by one point, but the fact that everything else was that close is the story. And yet, fans will say that a team that went 14 of 20 from the line and lost by 3 points "left six points at the line," because a kid who ran hit butt off and came up with five steals to keep his team in the game was dead tired and fell short on a couple of free throws. At 14 of 20 you have hit what it typical. If you go 16 of 20 that is a PLUS TWO over what you should have scored. The fact is FTA/FGA is the least significant of the Four Factors, and the way to make that state more meaningful is the make it FTM/FGA - that adds the difference in being able to get to the line AND the ability to hit once there to make one more significant stat instead of two minor ones.

Deadball turnovers by a team's offense. Finally, people overvalue how much dead ball turnovers hurt a team. They are bad. they are worse than a missed shot because you do not have the one in three chance of getting an offensive rebound and still trying to score. But giving up steals is so much more damaging because it can make a solid defense irrelevant and get an opponent who is struggling shooting on track with a couple of breakaway dunks.

To illustrate this I took Pomeroy's stats on offensive steals allowed and overall turnovers allowed, and subtracted to see how many turnovers were deadball (traveling, ball thrown out of bounds, 30-second violation).

Admittedly cherry picking a bit, I pulled out the 13 team for which well over half of their turnovers are LIVE BALL turnovers (first 13 listed), and the 13 teams that give up mostly DEAD BALL turnovers (last 13 listed).

UCLA overall ranks as the 17th best defense in the country even though 56% of their turnovers are live ball, and I believe it is a really bad Achilles Heel for what many believe is the best offense in the country but is starting to struggle. Overall these 13 teams give up slightly higher than average turnovers (19.2), but most of these are live ball steals (10.8 or 56%) and despite UCLA's nice defensive rating of 17th, these 13 average only the 194th best defense in the country.

The 13 at the other end of the spectrum give up almost as many turnovers (18.5), but only 6.7 are steals compared to the 10.8 by the first 13 teams. Overall their defenses rank 123rd on average, 71 spots better than the high steals allowed teams. Cincinnati from this group ranks only one spot better than UCLA overall as the 16th best defense in the country, but the fact that two-thirds of their turnovers are dead ball and only 5.4 per 100 trips are live steals means come tournament time you are going to need to grind it out against their half court defense to win.

Turnovers are not all equal, and looking at the steals allowed is a better indicator of vulnerability that simply looking at turnovers allowed - one of the four factors.


Worst % Live Ball TOConfTO%Stl%Deadball%StealDef Rank
Miami OHMAC19.411.4859%276
New MexicoMWC18.510.77.858%111
UMass LowellAE21.812.69.258%281
DavidsonA1017.710.27.558%80
Oral RobertsSum179.77.357%213
IowaB1018.310.47.957%91
UCLAP12168.97.156%17
CampbellBSth18.610.38.355%296
DrakeMVC18.810.48.455%227
NavyPat22.412.210.254%174
VMISC19.910.89.154%315
High PointBSth20.711.29.554%265
Ball St.MAC21.111.49.754%172
Ave. High Live Ball%19.210.88.556%193.7
Best % Live Ball TOConfTO%Stl%Deadball%StealDef Rank
Ohio St.B1018.87.111.738%63
Cal St. BakersfieldWAC21.78.113.637%120
Santa ClaraWCC16.2610.237%167
Mississippi St.SEC17.96.611.337%89
Cal St. FullertonBW21.8813.837%311
Cleveland St.Horz20.37.412.936%232
CaliforniaP1218.76.811.936%56
New Mexico St.WAC20.47.41336%85
Morehead St.OVC18.86.81236%185
Weber St.BSky17.1611.135%140
DaytonA1017.2611.235%36
CincinnatiAmer15.75.410.334%16
College of CharlestonCAA16.25.111.131%94
Ave. Low Live Ball%18.56.711.936%122.6

Value Add > Vegas on Point Value of Injured or Returning Player

The explanation of how many points a college basketball player is really worth to his team is pointed out in this post, which shows the video of a Las Vegas odds maker incorrectly assuming that a college player is never worth more than two points. Another writer then took this information and published a very safe projection on Fresno State staying closer to Oregon than Vegas predicted for a winning bet recommendation.

The interview from an otherwise very astute oddsmaker is here.

September 9 Prediction Email: "Clinton 3% Win = President Trump based on WaPost 50-state poll"

EMBARGOED - this is a background link for a reporter. Do NOT provide this link without permission unless and until this line is removed.

On September 9, virtually every media outlet in the country and thousands of others received an email from me correctly predicting the election result with a subject line,  "Clinton 3% Win = President Trump based on WaPost 50-state poll." No one believed my analysis, but many remembered when two months later I was within a point on the popular vote (2.1% Clinton) and almost perfect on the Electoral Map (I gave Trump Colorado but had him just short in Pennsylvania).

Sabato, Larry J.  11/9/16

11/9/16
to me
John,

You were right all along. I was wrong. Send crow extract. :) Congratulations.

Larry

Dr. Larry J. Sabato
Director, Center for Politics
University Professor of Politics
University of Virginia
So how in the world were analytics people at the Huffington Post and elsewhere so blind to what the numbers clearly showed was happening by Labor Day? Sports and political fans both like to attack "Moneyball" type predictors, as I was called on Fox News.

The only two analysis I know of who really were close were Nate Persily at Stanford, with Nate Silver at least in the ballpark.

Many people at events in San Francisco, New York and Washington DC tell me I am only one who predicted a Trump win in writing to them using actual analytics. They are all among the thousands of media and opinion leaders who remember the September 9, 2016 email below pointing out we were on a likely course to a 3% popular vote win by Hillary Clinton, but a 291-247 Electoral vote by Donald Trump (see below, with figures highlighted).

Among the many people I respect who are often correct when I am wrong, perhaps the highest complement came from someone with whom I compared notes throughout the campaign, the great Larry Sabato with the email above.

When some in Athletic Departments complain about one of their players not being in the top 100 (of 4000 players) at  www.valueaddbasketball.com, I often wish there were some way to prove to them their favorite players high turnovers, bad defense and missed shots really do make them less valuable than other players who score less but avoid mistakes. It is a funny intersection of sports and politics, as one of my favorite ESPN writers was kind enough to note in this post:

When you think about the intersection of advanced sports statistics and political forecasting, you think of Nate Silver. Far fewer people will know the name John Pudner, but college hoops fans probably should.

The fact is many fans of sports and politics have two things in common, the game looks simple so they believe they would be a better coach/campaign manager than the professionals, and they simply can not put aside their emotions and just look at the numbers objectively.

In the case of the election it was really not hard to figure out why Trump would likely pull within three points as predicted in my email two months out (he actually did even better, pulling to just a 2.1% loss), and in turn become President with about a 291-247 Electoral College margin (he actually won 304-227 by claiming Pennsylvania, which you see from the email below noted that he would get if he could pull within 2.5% in the popular vote).

The two obvious facts:

1. Trump was finally raising enough money to advertise in the closing two months ("Trump Shatters GOP records with small donors" was written by Politico 10 days after my email, but word was already out on the scale of his fundraising.)

2. Trump had a message that would move millions of voters once he finally had the money to get that message out (more than 37,000 shared my July 22 post titled "Donald Trump's Speech Grabs 14.8 Million Extra Likely Voters").

3. If you have the money and message to move the popular vote a couple of points, most states are going to move the same amount in your direction and the very numbers the Washington Post came up was fantastic research - but given the first two items it showed a likely Trump victory.

Nate Silver followed the logic of a Trump surge lifting all states flawlessly, and knew he had a chance. The only thing he was missing in giving Trump a one in three chance was #2 above, the power of his message.

With the Huffington Post giving Trump a 2% chance, me calling him the likely winner two months out proved more memorable than the email referenced in Taxation Only With Representation to 93,000 people telling them Dave Brat was in the process of upsetting Congressional Majority Leader Eric Cantor. That email will certainly be more remembered than the only other book with my name on the cover, a book with the stats of every Marquette basketball player in history that I see you can now buy on Amazon for anywhere from 10 cents to $2108 on Amazon (I'd take the 10 cents copy, even though it may not be in as good condition LOL).

Here is a copy of the email sent on September 9 pointing out that the Washington Post poll itself showed Trump should win the election. I added emphasis on the 291-247 electoral map win.

From: John Pudner [mailto:pudnerjohn@gmail.com]
Sent: Friday, September 09, 2016 7:29 AM
To: (sent to thousands of media and leaders)
Subject: Clinton 3% Win = President Trump based on WaPost 50-state poll


I call people all day about Take Back Our Republic’s conservative solutions for campaign finance, but knowing my history in number crunching (see Fox NewsCNN/Sporting News) almost everyone wants my take on the Presidential race first. Let me get it out of my way – forgetting my personal preferences on the race I put the results of the Washington Post 50-state poll on the attached table and the numbers lead me to two conclusions:

1.       Hillary Clinton has a MUCH better chance than Donald Trump at an Electoral Vote blowout, but …
2.       Trump actually has a slightly better chance than Clinton to win the Presidency.
I ran these numbers right after the Post poll came out, but I’ve only shared the basic math with select friends over the several days. The feedback was that those in DC who crunch numbers because their business models are heavily impacted by each President have models that are very close to this, and a number of Hollywood affiliated people who were told these numbers started making plans to leave the country. But seriously, below is the state by state and key notes, and I welcome any feedback or rebuttal, from any of the best predictors in the country from Larry Sabato at the University of Virginia to Nate Silver.

A.      The first thing you need to look at in any poll is when it was run. The Washington Post poll was run April 9 to September 1 – and all but the last three of those days the RealClear Politics average had Clinton ahead at least 5.4%. And in fact, when I weighted each Washington Post State Poll by the state population, the total of the Washington Post Poll was exactly a 5.4% lead for Clinton nationally – so it looks very accurate for when it was taken.

B.      If the poll did reflect the Election Day results, then Clinton wins by 5.4% and wins a blowout 375-163 Electoral College win as she puts Arizona, Texas and North Carolina (I just split Georgia and North Carolina since they were both tied in the WaPost poll). So a Clinton blowout is more likely than a Trump blowout.

C.      However, if Trump improved just 2.5% across the board, he still loses the popular vote by 3%, but that 2.5% improvement delivers North Carolina, Arizona, Texas, Colorado, Florida, Michigan and Wisconsin and Trump wins 291-247.  Right now Trump is within 3% nationally, so with the map really favors him in a close race.

D.      If Trump won the popular vote, or even just tied, then he would only go a little higher – a 317-221 Electoral win – so it is harder for him to get to a blowout win even though he seems to have an ever so slight advantage to win the race.

Keep the polling dates in mind when looking at state polls. An Arizona poll was released separately showing Clinton up one point there and was referred to as evidence she had a huge lead, but actually that poll was also run almost entirely when she was up 6 points or so nationally (August 17-31) and had the same result as the Post poll - a one-point Arizona lead. The real takeaway from an older poll like that is that Trump is running 5 points better in Arizona than he is nationally, so If Clinton wins by 7 points she likely does take Arizona along with Texas and Georgia. But as indicated above, if Trump is even within a few points nationally he wins the Presidency assuming a 2.5% shift in all the state polls the Washington Post ran to give Clinton the +3% popular vote win and Trump the Electoral win.

Of course not every state will shift the exact same amount, so if it were a 3% Clinton win it is likely that she would get at least one of the projected narrow Trump wins in Colorado, Florida, Michigan and Wisconsin – but unless that one state was Florida Trump would still be above 270 Electoral votes.

Here are all 50 states in the Washington Post poll shifted 2.5% toward Trump to account for a projected 3-point Clinton win in the popular vote. Please email back with other perspectives or anything you see is off.

The states are listed from strongest Trump to strongest Clinton, and how many Electoral votes Trump would have if he won only that state and the states before it. The percent margin is how much Trump would win (positive number) or lose (negative number) the state by assuming he improved just 2.5% in each state from the Post Poll – to estimate the result if Clinton won the popular vote by 3%. So if Trump lost by 35 points, he would only win Wyoming to lose 3-535 in the Electoral College. If he won by 35% he would win every state but lose DC to lose 535-3, etc. You need 270 to become President.

Wyoming         40.5%3
North Dakota 30.5%6
Oklahoma       26.5%13
West Virginia  26.5%18
Kentucky         25.5%26
Alabama          23.5%35
Idaho21.5%39
Tennessee      20.5%50
Indiana 17.5%61
Louisiana         17.5%69
South Dakota 16.5%72
Arkansas          15.5%78
Montana         15.5%81
Kansas  14.5%84
Utah  13.5%90
Nebraska         13.5%95
Missouri           12.5%105
Alaska   10.5%108
South Carolina   9.5%  117
If Clinton were to improve 4.5% from the WaPost Poll, she wins by 10% and takes Electoral College 421-117                  
Iowa  6.5%  123
Ohio  5.5%  141
Mississippi      4.5%  147
Georgia 2.5%  163
When weighting the WaPost 50-state by state population, Clinton wins popular vote by 5.4%, and 375-163 Electoral College Win                    
North Carolina   2.5%  178
Arizona 1.5%  189
Texas1.5%  227
Colorado          0.5%  236
Florida  0.5%  265
Michigan          0.5%  281 (Trump needs to get to here to win)
Wisconsin        0.5%  291

If all States in the Washington Post poll shifted 2.5% toward Trump, then Clinton wins popular vote by 3%, Trump wins 291-247 Electoral                         

Pennsylvania -1.5% 311
Nevada -2.5% 317
If Trump and Clinton were to tie, meaning a 5.4% shift toward Trump since the Post polled Aug. 9-Sept 1, then Trump wins 317-221                  
Maine   -5.5% 321
Virginia -5.5% 334
Minnesota      -6.5% 344
New Hampshire -6.5% 348
Rhode Island  -7.5% 352
Connecticut    -9.5% 359
Delaware         -11.5%  362
New Mexico   -11.5%  367
Illinois   -12.5%  387
New Jersey     -12.5%  401
Washington    -13.5%  413
Oregon -16.5%  420
New York         -19.5%  449
Massachusetts  -20.5%  460
California         -21.5%  515
Vermont          -25.5%  518
Hawaii   -27.5%  522
Maryland         -27.5%  532
District of Columbia (not actually polled by WaPost, but will be Clinton)    -67.5%  535

Sunday, January 29, 2017

Value Add Version 5.0 Corrects for Trips Per Trip, Domination Distortion

EMBARGOED - THIS IS A BACKGROUND PIECE FOR MEDIA REPORTERS, AND CANNOT BE LINKED AS LONG AS THIS LINE IS HERE OR UNTIL PERMISSION IS GRANTED.

Version 3.0 of Value Add corrected slight variations to ensure the sum of their players' Value Adds equalled the offensive and defensive team ratings on www.kenpom.com. If a team has an offensive rating of 108 and the sum of the players' Value Add indicates they only add up to a 106 offensive rating, then each player is adjusted upward precisely to account for the fact that they are doing something that is not showing up on the stat sheet. Likewise if the team is underperforming what their stats would calculate, each player is adjusted downward.

In January 2017, Value Add 5.0 calculated for two additional factors to adjust the 3.0 ratings:

1. Version 3.0 indicates the players Value Add in terms of how much each player impacts a team's trips up and down the court per 100 trips down the court. However, fans typically want to know how many points a player is worth to their team PER GAME. I have had callers tell me that even the great Ken Pomeroy cannot seem to explain the PER 100 TRIPS vs. PER GAME distinction in sound bites on the radio - as Jeff Sagarin's basketball rankings are much easier to understand as a "points per game" measurement.

We start by looking at www.kenpom.com to calculate that the average team during the tempo free era has had 67.2 trips down the court (possessions) per game - so the basic adjustment to get a POINTS PER GAME Value Add is to multiply their Version 3.0 Value by .672 to translate the final result to match up with Sagarin's rankings.


YearAve Trips
200269.5
200368.5
200467.7
200567.3
200667.0
200766.9
200866.9
200966.5
201067.3
201166.7
201266.1
201365.9
201466.4
201564.8
201669.0
201769.3
Average   67.2

2. The second factor adjusted by Version 5.0 is the distortion to the Value Add rating of players on great teams who sit on the bench when their teams are blowing out lesser opponents. These games are particularly damaging to a rating, because not only do the key players not get as many stats in those blowouts, but because they are often against teams with poor defenses, they receive a lower rating for those games due to going up against an inferior defense.

In going through the sixteen seasons, it appears the good players on the teams that win by the biggest margins lose enough Value Add to justify a multiplier of 1.3 to account for the extra minutes and possessions they could have had against the lesser early season opponents.

On the flip side the stars on poor teams often play almost all the minutes while getting blown out by the great teams, getting credit for scoring against the greatest defenses in the land even though points often come against the subs and even walk-ons for the great team. A review of the seasons indicates that the team that gets blow out by the biggest margins should have their Version 3.0 ratings multiplied by 0.695 to account for the easy points and other stats they get.

Each team gets a different multiplier, with the team winning by the second biggest margin getting a slightly smaller multiplier than the 1.305, all the way down to the players on the 351st team getting a 0.695.

However, we accomplish both items 1 and 2 in this piece by also multiplying those values by the .672 to account for trips per game, so the top victory margin team actually has their Version 3.0 Values multiplied by 0.877, and the team with the second biggest margin a slightly lower factor, making the actual factor for the 351st team 0.467.


Adjustment from v3.0 to v5.0    Domination Distortion    Multiplied by Trips/Game
Teams with biggest ave.victory margin     1.305    0.877
Average Team     1.000    0.672
Team with biggest average defeat margin     0.695    0.467

Some pure analytics people who attend MIT Sloan do not like any factor that is not purely mathmatically based, but I subscribe to the Bill James' approach that when observation gets you very close to an accurate mathematical equation, you administer it in a way that gets you as close as possible to the truth.

A player's Version 5.0 Value Add Ranking indicates how many points he can be worth to a team. However, the next question is how many points the player is worth over the replacement players of the actual team on which he plays.

When Version 5.0 was first calculated January 29, 2017, the average Value Add of the top player on each of the 351 teams was 5.13. The following is the average of each of top nine players (so the average 9th player of all 351 teams is 0.24.


TeamAve Value Add
Top Player5.13
2nd Player3.58
3rd Player2.69
4th Player2.02
5th Player1.46
6th Player1.01
7th Player (Subtract)0.68
8th Player0.42
9th Player0.24

This is the rating on which each player is ranked. The true value of a player to the actual team on which he played is usually lower, as it is approximately his Value Add minus the 7th best player on his team. So on a typical team:

The actual point value of the best player on a team is actually 5.13 MINUS the 7th best players total of 0.68 for a true value of 4.45.

The fact is that losing a player that is not one of the top six players is tiny and not worth adjusting the projected result of an upcoming game. Certainly there can be a starting point guard who has a low Value Add rating because he misses so many shots and turns the ball over so much that his Value Add is low, but in general it is very low.

For this reason, in the initial Version 5.0 Ratings on January 29, 2017, Gonzaga's Nigel Williams-Goss was the 20th most valuable player on Gonzaga's team at 9.21 points per game. He would in fact be worth 8.53 points per game to a typical team - so could turn a five point loss into a three- or four-point win, and that is the standard we use to measure his value against other teams.

However, the Gonzaga 2017 team is so loaded that Williams-Goss is actually only worth 4.91 points per game. Losing five points per game off your Sagarin rating is huge when you are competing for a national championship. When these figures were calculated, Gonzaga's Sagarin rating was 93.92 - second in the country - and subtracting 4.91 leaves a "Gonzaga without Williams-Goss" Sagarin Rating of 89.01 for 16th in the nation (Gonzaga/Sagarin 93.92 MINUS losing Williams-Goss 9.21 PLUS 4.45 for Gonzaga's 7th player equals 89.01.

Domino Effect, Not Baseball War

Why? Because his possessions and minutes are being used by players who are almost as good. Gonzaga's Josh Perkins has the best Value Add of any 7th man in the country at 4.30, so the domino effect from Williams-Perkins being out for a game to everyone down to Perkins getting more opportunities calculates as 9.21 (Williams-Goss Value Add) MINUS 4.91 (Perkins' Value Add) to equal 4.91 (actual point shift in the final score by Williams-Goss missing a game.

Many call me and try to figure out the Value Add of the backup player who will most likely replace the injured player, but unlike Baseball WAR where a specific player replaces another specific player both in the lineup and at a position in the field, the possessions are spread among several players and positions are even moved throughout the game to compensate. The one exception is when a great point guard is injured and the team lacks a decent backup point guard.

Louisville has the second best 7th man, and the following are the 40 teams with such a strong top seven that you need to subtract at least 1.70 from any of their player's Value Add ratings to get the true number of points the team will be hurt if the player is out.


RnkBest 7th Players 1/29/2017Teams w/ Deepest BenchSubtract from Value
1Josh Perkins^ 13Gonzaga4.30
2Anas Mahmoud^ 14Louisville4.10
3Justin Leon^ 23Florida3.51
4Jack Salt^ 33Virginia3.48
5Shaquille Morris^ 24Wichita St.3.44
6Nate Fowler^ 51Butler3.44
7Casey Benson^ 2Oregon3.03
8Terry Maston^ 31Baylor2.87
9Nate Britt^ 0North Carolina2.71
10Lamont West^ 15West Virginia2.71
11Carlton Bragg^ 15Kansas2.60
12Khalil Iverson^ 21Wisconsin2.60
13Muhammad-Ali Abdur-Rahkman^ 12Michigan2.49
14Mychal Mulder^ 11Kentucky2.45
15Katin Reinhardt^ 22Marquette2.43
16Ryan Cline^ 14Purdue2.33
17Temple Gibbs^ 2Notre Dame2.29
18Devontavius Payne^ 1East Tennessee St.2.27
19Eric Paschall^ 4Villanova2.27
20Jarquez Smith^ 23Florida St.2.26
21Niem Stevenson^ 10Texas Tech2.24
22Chance Comanche^ 21Arizona2.24
23Nathan Taphorn^ 32Northwestern2.15
24Jordan Murphy^ 3Minnesota2.14
25Alvin Ellis^ 3Michigan St.2.08
26Sedrick Barefield^ 2Utah2.08
27Shembari Phillips^ 25Tennessee2.06
28Trey Thompson^ 1Arkansas2.04
29Kamari Murphy^ 21Miami FL2.01
30Terrence Samuel^ 5Penn St.1.97
31Isaiah Zierden^ 21Creighton1.96
32Tre Scott^ 13Cincinnati1.94
33Drew Urquhart^ 25Vermont1.94
34Sam Miller^ 2Dayton1.92
35Paul Watson^ 3Fresno St.1.91
36Jarvis Garrett^ 1Rhode Island1.89
37Tyrique Jones^ 0Xavier1.89
38Carlbe Ervin^ 1Kansas St.1.75
39Horace Spencer^ 0Auburn1.70
40Jonah Mathews^ 2USC1.70

The following are the original top 25 players calculated using Version 5.0 based on the Louisville, Michigan State and Villanova wins on January 29, 2017. The entire 4000+ players were released at www.valueaddbasketball.com shortly thereafter:


RnkPlayerPts/Gm>ReplTeamConfHtClass
1Josh Hart^ 313.03VillanovaBE6'5"Sr
2Frank Mason^ 012.35KansasB125'11"Sr
3Luke Kennard^ 511.98DukeACC6'6"So
4Monte Morris^ 1111.37Iowa St.B126'3"Sr
5Lauri Markkanen^ 1011.15ArizonaP127'0Fr
6Yante Maten^ 110.41GeorgiaSEC6'8"Jr
7Ethan Happ^ 2210.39WisconsinB106'10"So
8Jevon Carter^ 210.35West VirginiaB126'2"Jr
9Ben Lammers^ 4410.13Georgia TechACC6'10"Jr
10Marcus Marshall^ 110.01NevadaMWC6'3"Sr
11Alec Peters^ 259.79ValparaisoHorz6'9"Sr
12Markelle Fultz^ 209.65WashingtonP126'4"Fr
13Jaylen Adams^ 109.58St. BonaventureA106'2"Jr
14Jalen Brunson^ 19.49VillanovaBE6'2"So
15Caleb Swanigan^ 509.48PurdueB106'9"So
16Sindarius Thornwell^ 09.32South CarolinaSEC6'5"Sr
17TJ Williams^ 109.32NortheasternCAA6'3"Sr
18Donovan Mitchell^ 459.29LouisvilleACC6'3"So
19Lucas Woodhouse^ 349.26Stony BrookAE6'3"Sr
20Nigel Williams-Goss^ 59.21GonzagaWCC6'3"Jr
21Jeffrey Carroll^ 309.20Oklahoma St.B126'6"Jr
22Dennis Smith^ 49.15North Carolina St.ACC6'3"Fr
23Bonzie Colson^ 359.05Notre DameACC6'5"Jr
24Joe Chealey^ 139.05College of CharlestonCAA6'4"Jr
25Jock Landale^ 349.05Saint Mary'sWCC6'11"Jr