<<

Predict NBA 2016-2017 Regular Season MVP Winner Mason Chen, Black Belt, Stanford OHS Who should win MVP Awards?

1 2015-2016 KIA NBA MVP Award Result

• Voted by 131 Sports Media Reporters (Cluster Sampling Method) • Curry got all 131 Votes in 2015-2016 Season (consensus choice) 2 Most Controversial MVPs in Sports (Questionable Cluster Sampling) 4. over , 1990 In the closest NBA vote since the media took over the voting in 1981, Johnson edged Barkley by 22 points, even though Barkley received more first-place votes (38 to 27). Barkley averaged 25.2 points and 11.5 rebounds and shot 60 percent from the field. Johnson averaged 22.3 points, 11.5 assists and 6.6 rebounds (while shooting 48 percent). Barkley's Sixers won 53 games while the Lakers won 63. Oh yeah -- averaged 33.6 points, 6.9 rebounds and 6.3 assists and finished third in the voting. 7. over , 2002 Duncan averaged 25.5 points and 12.7 rebounds for the Spurs while Kidd engineered an amazing turnaround for the Nets after coming over from Phoenix: from 26 wins to 52 and a trip to the NBA Finals. He was attempting to become the first guard since Magic Johnson in 1990 to be named MVP, but Duncan's scoring won out over Kidd's playmaking. Kidd had averaged 14.7 points, 9.9 assists and 7.3 rebounds, but shot just 39.1 percent (no MVP had shot that poorly since in 1957). 10. over Michael Jordan, 1997 Jordan had already won four MVP Awards, so maybe voters were simply getting a little tired of giving it to him. Jordan had another typical MJ year: 29.6 points, 5.9 rebounds and 4.3 assists. The Bulls, however, did slip from a record 72 wins to 69. Malone averaged 27.4 points, 9.9 rebounds and 4.5 assists in leading Utah to 64 wins. Malone topped Jordan in a close vote (986 points to 957), but Jordan would get the last laugh: the Bulls beat the Jazz in six games in the NBA Finals.

http://www.espn.com/page2/s/list/MVPcontroversy.html 3 Build Predictive Model (Function)

Actual Predict 2003-2016 2016-2017 MVP Winners MVP Winners (Y) (Y)

2003-2016 2016-2017 Player Statistics MVP Model Player Statistics (X) Y= F(X) (X)

2003-2016 2016-2017 Team Record Team Record (X) (X)

• Use historical data (2003-2016) to build the MVP Model • Use 2016-2017 player statistics to predict the new MVP Winners 4 Add Four Combining Statistics (Interaction Effect, Trim Insignificant Terms)

Y (MVP Index, Ranking) = Function (Player Statistics, Team Record) Authors have taken SME opinions and created four Combining Statistics: (1) RB/MIN, (2) AST/MIN, (3) A/T Ratio, (4) PPG/MIN which can demonstrate both offense & defense efficiency 5 Data Transformation to Z-Scale N(0,1)

Y (MVP Index, Ranking) = Function (Player Statistics, Team Record) • Z-Transformation will transform each individual players’ statistics category to the normal Z scale for direct comparison (eliminate sample mean and sample standard deviation bias) 6 Create MVP Index and Uniform Model

Y (MVP Index, Ranking) = Uniform Function (Player Statistics, Team Record) • Create MVP Index: by summing each Z-score of all statistics categories (Uniform Model) 7 Evaluate the Uniform Model

Some correlations based on Uniform Model, but not very strong prediction between MVP Index (prediction) and Actual MVP Rank 8 Root Cause Analysis of Uniform Model

Not Significant

Certain Significance

• Some player Statistics are more relevant than the others on MVP Rank (How to Quantify Significance?) 9 Derive Weighted Model Actual • Use descriptive statistics Variable Rank Median (median, range) of top two GP-N 1 0.712 2 0.635 ranked players vs. non- 3 0.443 4 0.405 ranked players to quantify 5 0.520 the Weight Factors * 0.3283

Top 5 Factors • Weight coefficients will be added to the Uniform Model 1. Point per becoming the Weighted 2. Point per Game Model 3. % 4. per Game 5. per Game 10 Best Subset (Feature Selection) Enhance the Weighted Model

Top 5 Variables: 1. Field Goal % 2. 3-Point % 3. Rebound per Game 4. Games Played (Healthy, Durant) 5. Point per Minute Healthy Durant was leading the MVP race before being injured in the mid season (Durant won Final MVP) 11 Compare two Weighted Methods

Weighted Coefficients Best Subset Selection 1. Point per Minute Inflated 1. Field Goal % Power 2. Point per Game VIF > 5 2. 3-Point % (Curry) 3. Field Goal % 3. Rebound per Game 4. Assist per Game 4. Games Played (Healthy) 5. Rebound per Game 5. Point per Minute

This Best Subset algorithm can minimize the Multi-Collinearity (shown in Mallows Cp coefficient) among the linearly dependent factors.

12 Model Accuracy: Uniform vs. Weighted

Weighted Model has slightly improved the model accuracy 13 Power Model (Consider Team Record)

• Both Uniform and Weighted Models considered Player Statistics only • MVP is the Most Valuable Player. Therefore, player’s contribution to the team record should be considered • Majority of the past MVP winners are from the Best or Better Teams • Model should set Team Record as a higher priority above the Individual Statistics Performance (Power > 1) • Created a new Power MVP Index= Weighted Index * (Team Winning%) ^ (Power) • Individual Weighted Model= Power Model (Power= 0) 14 Collect NBA Team Record

Y (MVP Index, Ranking) = Function (Player Statistics, Team Record) http://www.basketball-reference.com/leagues/NBA_2016_standings.html 15 Optimum Power Model (Prevent Over-fit)

Player Statistics ~ Team Record

Power= ∞ Power= 2 Let the best team decide Optimum their own MVP (even more subjective)?

Weighted

Power= 2 model has shown 67% Prediction Accuracy 16 2016-2017 Regular Season MVP

Around mid April after completed the regular season, we have predict MVP winners two months earlier than official NBA announcement 17 Co-MVP (’s Suggestion)?

Team Performance is more Individual Performance is important (Power >1.5) more important (Power < 1.5)

The Winner is Westbrook (released on June 26, 9PM ET, 2017) 18