Вы находитесь на странице: 1из 54

Click icon to add picture

PLAYER
EFFECTIVENESS-2
By Group 5:
Nelisent; Fashid; Ayush; Gaurav; Divya

Agenda
Counterpoints
Quality vs Quantity
Finding the open receiver
WLW Concepts
The four factor model
Linear weights for evaluating NBA players
Adjusted Player Ratings

COUNTERPOINTS
Advanced defensive metrics for NBA basketball

Introduction
What is the paper about?
Objective of basketball is to score points on offense and to prevent points
on defense
Games analytics only evaluate offensive performance while defensive
performance is overlooked
This paper is designed to give a mathematical approach to measure the
defensive play in a basketball game

Terms
Volume score: The total magnitude of attempts which an individual
defender faces
Disruption score: The degree to which an individual defender is able to
reduce the effectiveness of his assignments shots
Defensive Shot Chats: Like shot charts, but for defensive play. Visual
depiction of an individuals defensive prowess; we map both volume and
score and disruption score across the scoring area
Shots Against: A weighted average of the shots attempted against the
defender per 100 possessions
Counterpoints: A weighted average of points scored against a particular
defender per 100 possessions

Method
Estimation of Defensive Matchups
Estimate who is guarding whom at any given moment
To identify information we estimate an average defender position as a
function of offender location, ball location and the hoop location
Modelled as:
Where m of a player at time k is taken, with Y0 + YB + YH=1 and O, B and
H representing the offender, ball and hop locations respectively.
Hidden Markov model to express the evolution of defensive matups over
the course of the possession

Matchup model
Define a defender volume score and defender disruption score
The volume score quantifies how often a defenders match up takes the shot when
guarded by this particular defender
The disruption score quantifies how much the defender reduces the opponents shot
efficiency
We follow the average strategy of the players and run a multinomial logistic
regression, where each outcome represents a shooter location pair. We include
matchups, and shooter and defender identities as predictors for the ultimate outcome
To compute the disruption score for the shot charts, we use a logistic regression to
predict makes and misses. Here, we use shooter and shot defender identities and
defender distance to predict expected efficiency.
The coefficient related to defender identities define the volume score and disruption
score in each regression and correspond to the change in odds of a shot taken or a
shot made, respectively.

Counterpoints
Volume and disruption scores give insights into the what and why of a
players defensive abilities , their primary limitation is that they are staticthey dont account for how the possession unfolds in time
A defender does not guard the same offender for an entire possession of
the ball
This notion is looked upon by computing a variation on the defenders
volume score and disruption score at each moment t seconds before the
time of the shot. Using this concept , we can identify how often a defenders
initial matchup eventually shoots or scores

The result of this method


The probability of a defenders matchup taking the shot varies as we look
back in time from the moment of the shot
Case study 1: Original Matchup: The original matchup looks at the first
can summarize defensive performance in the early possession regime by
isolating seconds 10 through 4 before a shot attempt. For each moment
before a shot, we examine two measurements the average number of
attempts taken and the average number of points scored against a
defender

Case study 2: Before Shot Matchup:


Another summary of a defensive players performance is by simply looking at all
matchups for a defensive player at the moment of the shot and compute the attempts
against and points against values
To do this , we look at all of the possessions played by a defender and count how many
times they were defending the shooter at the moment of the shot and how many points
were scored per 100 possessions
Case Study 3: Fractional Method:
Both of these approaches yield interesting insights about defenders strengths and
weaknesses. But , both approaches have flaws
The pre shot matchup disproportionately penalizes bigs who contests more shots
On the other hand the early shot matchup method may ignore the responsibility of the
defender who contests the shot

Key Findings
Despite the player tracking revolution in basketball, assessments of
performance have been heavily biased toward offensive play
Defensive ability is difficult to quantify with a single value
Summaries of points scored against and shots attempted against can say
more about the teams defensive scheme rather that the individual players
defensive ability

QUALITY VS QUANTITY
Improved Shot Prediction in Soccer using Strategic features from spatiotemporal data

Introduction
2014 FIFA World Cup in Brazil,
Germany blitzed Brazil in the semi-final 7-1.

Total Shots
Shots on
Target

Brazil
18
13

Germany
14
12

Match statistics does not reflect the sheer dominance that Germany had
Paper addresses the question How can we quantify the value of a shot
directly from player tracking data?
High Probability - Low probability

Goal: Estimate the shot quality from tracking data


Data: Prozone tracking data. Nearly 9732 shots from a season(10 sec
window)
20 teams A-T
353 games
Ten-second window before a shot of 9732 shots

Expected Goal Value (EGV)


Expected Goal Value: Given a shot, what is the probability it will
end up a goal?
Approximately 9.6 % of all shots end up in a goal (EGV= 9.6%)
Average Error = 0.1745
Other match contexts
1. open-play (possession in the forward third)
2. counter-attack (players break quickly from one-end to the
other)
3. Corners
4. Penalties
5. free-kick (shot on goal from a free-kick)
6. set-pieces (a cross that comes into the box from a free-kick)

Team more likely to score on a


counter attack compared to
normal possession
Team more likely to score in
normal possession than a free
kick

Corner

Penalty

Error: Incorporating
Match Context

Free Kick

Incorporating Location

All Shots

Error: Incorporating
Location
+ Match
Context

All Goals

Defender Proximity

Goal Side
(GS)

Non Goal-Side
(NGS)

Game Context EGV Goal Side

EGV Non
Goal-Side

Open Play

7.49 %

11.59 %

Counter Attacks

12.46 %

18.44 %

Other Features: Actions & Team Motions

Actions of the players plays an important


role as well
Whether the player passes, dribbled
Goal keeper location
Team dynamics- the team motion

Team Efficient Ratings: All


Contexts

Offensive

Defensive

Open Play

Offensive

Defensive

Match Analysis

Example 1

Example 2

Example 3

FINDING THE OPEN RECEIVER


An quantitative geospatial analysis of quarterback decision- making

Motivation
https://www.youtube.com/watch?v=IBC
Nmp9WHKs
RFID technology monitors on-field location
of players
How QB decisions and passing ability
impact the likelihood of receiver
concussions
By making better decisions and finding
the open receiver, QBs can put their
receivers and their teams in better
positions to succeed

NFL QB Passer Rating vs. Concussions and


Team Losses for three years

Data Collection
Captured spatial coordinates of all 22 on-field players based on game video
at 3 Hz to the nearest 0.25 yard for every base offensive pass play
Success outcomes: completions; yards gained.
This dataset focuses on 1st and 2nd down plays because the marginal team
benefit of additional yardage is more direct, compared to 3rd and 4th
downs
A significant benefit is earned when an offensive team achieves more
yardage than the yard to gain on 3rd and 4th downs
Down and Distance Utility Function

Receiver Openness
The instantaneous player velocity is important in addition to positional data,
a predictive tessellation (Voronoi Tessellation) is developed to quantify
receiver openness
Zone Size: How much of the field a receiver owns. By taking each
players instantaneous positions and velocities, the tessellation is
performed after projecting each players position forward two frames (2/3 of
a second) and finding the respective zone area

Zone area distribution for every


eligible receiver at QB decision point

Zone Integrity: Projected distance to the nearest defender from an eligible


receiver
Openness Classification:
Wide-open: ZA > 200 yards2; I > 8 yards
Open: ZA between 100-200 yards2; I between 4-8 yards
Defended: ZA between 50-100 yards2 ; I between 2-4 yards
Well-defended: ZA < 50 yards2; I < 2 yards
Hence, the QB completed a higher percentage of passes to open and wideopen receivers than those who were defended or well-defended
Completion percentage as a
function of targeted receiver
openness

Expected Gain (Yardage): Observing the maximum y-value of a receivers predicted zone.
The receivers maximum gain would occur at the point in his zone which is the furthest point
possible down field.
A poorly executed pass which forces a receiver to dive will reduce the yardage gained.
Alternatively, several broken tackles after the catch will lead to higher than expected actual
gain

Comparison of the expected gain for


every completed pass (total 148)
against the plays actual gain.

Player Elusiveness: The additional yardage gained by a player from the expected
yardage. The more yards a player gains than expected, the more elusive he is
Inference: Running backs (RB) are more elusive than Tight Ends (TE) and Wide
Receivers (WR)
The actions of the QB, receiver, and defenders ultimately define the actual yardage
achieved.

Receiver elusiveness for


148 completions

QB Decision Analysis
A QB can check down his options and decide for whom to target with his pass.
Each receiver frame is assigned an openness (wide-open, open, defended, and well-defended) and an
expected yardage of gain
The zone size and integrity are combined to quantify a receivers openness along with his expected gain
The expected payoff is:
P(CMP% for openness factor) * Expected Yards = Expected Payoff Yds
Percentile of the target receiver expected payoff: The expected payoff yardage is captured for all options
prior to the target selection, and the target payoff is compared to the play population options
80 percentile: ideal target decision
50 percentile: preferred target decision
20 percentile: neutral target decision
Below 20 percentile : undesirable target decision

QBs Decision Analysis

Summary and Future Work


Analyzing multiple QB and team strategies can help identify how QB tendencies affect
completion percentage and propensity for target decisions.
This study attempts to show the types of analysis possible with geospatial data available in NFL
The geospatial tracking allows for significant opportunities as the data is analyzed, so NFL and
American Football can cover the lag when it comes to the analytical player and team analysis
Limitations:
It projects instantaneous velocity as a constant (no acceleration); it assumes all players are
assumed to perform at their current velocities (no individual acceleration profiles).
QB decision-making does not account for receiver distance from line of scrimmage in the
completion probability equation
The data is only accurate to within 0.25 yards at 3 Hz. It is expected that the existing RFID
tracking technology provides more accurate, more frequent data points for analysis.

WLW CONCEPTS

The four factor model


Linear weights for evaluating NBA players
Adjusted Player Ratings

THE FOURFACTOR MODEL

Effective Field Goal Percentage


EFG = (all field goals made + 0.5 (3- point field goals made)/(all
field goal attempts).
EFG gives 50% more credit for making a three- pointer because a
three- pointer is worth 50% more points than a two- point field
goal

How Can We Evaluate Team


Rebounding?

The percentage of rebounds a team gets when they are on


offense called Offensive Rebounding Percentage, or ORP
The percentage of rebounds a team gets when they are on
defense (called Defensive Rebounding Percentage, or DRP

Four Factor Model:


Offense/Defense

Team Offense

Team Defense

Effective Field Goal Percentage (EFG)


Turnovers Committed per Possession
(TPP)
Offensive Rebounding Percentage
(ORP)

Opponents Effective Field Goal


Percentage (OEFG)
Defensive Turnovers Caused per
Possession (DTPP)
Defensive Rebounding Percentage
(DRP)

Free Throw Rate (FTR)

Opponents Free Throw Rate


(OFTR)

The Four- Factor analysis of the NBA 2006-07 season

The Four- Factors are virtually


Uncorrelated
Notice that most of the correlations are near 0: The interesting thing about the four
factors is that there is little correlation among them
Lets see the three largest (in absolute value) correlations:
There is a -.67 correlation between defensive shooting percentage and defensive
rebounding.
There is a -.47 correlation between offensive shooting and offensive rebounding
There is a .46 correlation between offensive rebounding and offensive turnovers.

How Important are the four factors?


Running a regression using the data of 2006-07
season can predict a teams number of wins from the
following four independent variables
EFG-OEFG
TPP-DTPP
ORP-DRP
FTR-OFTR
Games won = 41.06+351.88(EFG-OEFG)
+333.06(TPP-DTPP)+130.61(ORP-DRP)+ 44.43(FTROFTR)
These four independent variables explain 91% of the
variation in the number of games won. The standard
error of 3.53 means we are 95% sure our predicted
wins will be within 2(3.53)=7.06 wins of the actual
number of wins.

Click icon to add picture

LINEAR WEIGHTS
FOR EVALUATING
NBA PLAYERS

Linear Weights For Evaluating NBA Players


NBA Efficiency Rating
Efficiency per game = (points per game) + (rebounds per game) + (assists per game) + (steals per
game) +shots per game) - (turnover per game) (missed FG per game) (missed FT per game)
NBA efficiency rating was created by Dave Heeren
This system simply says that all good statistics are worth +1 and all bad statistics worth -1
This does not make sense why??
A player who shoots 26.67% on three- point field goals would raise his efficiency by taking
more three- pointers
Certain players were vastly overrated according to the NBA Efficiency statistic and other
players were vastly underrated in NBA Efficiency statistics of 2006-07 season

Player efficiency Rating system and Game Score Rating


An average NBA player has a PER score of 15
PER rating formula is complex, but as pointed out by Berri, a player who shoots more than
30.4% on two- point field goals can increase his PER rating by taking more shots
A player who shoots more than 21.4% on three- point field goals can increase his PER rating by
taking more shots
Implication: If you are by far the worst shooter in the league, you can help your team by taking
more shots

Game score = (points * 1.0)+(FGM *0.4)+(FGA *0.7)+(FTA-FTM) * 0.4)+


(OREB0.7)+(DREB
*0.3)+(STL *1.0)+(AST *0.7)+(BLK *0.7)+(PF *0.4)+(TO *1.0)

A player shooting over 20.4% (0.7/3.4) on three- pointers will increase his Game Score by
taking more shots
A player shooting over 29.2% on two pointers (0.7/2.4) will also increase his Game Score by
taking more shots

Wins scored and wins produced


Player wins scores = points + rebounds + steals + 0.5(assists)+ 0.5(blocked shots)
- FG attempts turnovers 0.5(FT attempts) 0.5(personal fouls)
Unlike the NBA Efficiency metric or Hollingers PER rating, the Linear Weights in the Win
Score metric seem much more sensible
For example, to raise his rating by shooting more, a player needs to shoot over 50% on
two- point field goals or over 33.33% on three- point field goals
It also seems reasonable to give equal weight to turnovers and rebounds, because a
rebound gets you possession and a turnover gives up possession
The nice thing about Wins Produced is that the sum of Wins Produced by all of a teams
players will be nearly the same as the teams total number of wins

For example, for the 20067 Chicago Bulls, BSB found the
Wins Produced shown in table 29.2
The total Wins Produced for the Bulls is 52.17, which is
close to their actual total of 55 wins
As an example, Bruce Bowen of the Spurs always fares
poorly in terms of Win Score and Wins Produced. Most NBA
experts agree that Bowen is a superb defender
For example, the box score does not give credit for the
following:
Taking a charge

Deflecting a pass

Boxing out so the


teammates gets
the rebound

The pass before


pass that earns a
assist

Helping out on
defense when the
teammates is
beaten by a quick
guard

Setting a screen
that leads to a
three pointer

ADJUSTED +/PLAYER RATING


Definition of a good Player:
Somebody who makes his team better; not a player who
scores 40 points per game

Pure +/- Ratings


Used in hockeys +/- statistic: How many goals a
players team outscores their opponents by when
the player is on the ice
Power plays (one team has more players) are excluded
from calculations

NBA players: pure +/- per 48 minutes on the court


is measured
Problem: Quality of the players he play with and
against is not counted
Ratings become useless
For example: Player A ratings is 0 against leagues
worst team; Player B rating is also 0 against a normal
team But A is a better player than B

Adjusted +/- : A Better Approach


Adjusted each players rating based on :
Ability of the player he is on the court with
Players he plays against

Data used to make this adjustment:


38000 rows of play-by-play data
Each row represents a segment of time during which
the players on the court remain unchanged

An average NBA player will have an adjusted


+/- rating of 0
Adjusted +/- rating of 5 : If the player
replaced an average NBA player for 48
minutes, his team would improve by an
average of 5 points per game
Example: Team 1 played 20 full length games
against Team 2
Team 1: 1 to 9 players
Team 2: 10 to 18 players

Score Prediction
The predicted score for each game:
= (sum of player ratings for Team 1
players in game)-(sum of player ratings
for Team 2 players in game)
Team 1 total ratings = -10.58 + 12.42
+ 1.43 - .58 + 2.43 = 5.1
Team 2 total ratings =7.42 + 1.43
+16.42- .57- 6.57 = 18.1.
Predict Team 1 to win by 5.1-18.1 =
-13 points

Calculation of Adjusted +/- ratings


The setup of the spreadsheet for determining Adjusted +/- ratings
In column B - begin by entering any trial set of Adjusted +/- ratings
These ratings are averaged in cell B3
In column S - Determine the sum of the Adjusted / for our team during each game
In column T - Determine the sum of the Adjusted / for our opponents during each
game
In column E - Created prediction for each game
=(sum of our player rating) [from column S]-(sum of opponent player ratings) [from column T]

In column D -Computed the squared error for each of our game predictions
= ((column G) (column E))2

Minimize SSE (in cell D2) by changing adjusted +/- ratings (cell B5:B22); constrain
the average rating (computed in cell B3) to equal 0

If the games are not of equal length


For each game you should predict the point differential per minute
Compute the squared prediction error for each game on a per- minute basis
To compute the Target Cell you weight each games per- minute squared
error by the number of minutes in the game

20062007 Adjusted / Ratings


Each players offensive ability and defensive ability have been included in
the ratings system
A good offense rating can be created by doing many things: scoring points,
getting rebounds, throwing good passes, setting screens, reducing
turnovers, and so forth
While a positive offense rating is good, a negative defense rating is good
A good defensive rating can be created by doing many things: blocking
shots, stopping the pick and roll, reducing turnovers, causing turnovers,
rebounding, and so forth
The beautiful thing about a WINVAL player points rating is that offensive
ability and defensive ability are weighted equally

There are several paths to basketball


excellence:
You can be great on both ends of the court
(e.g., Kevin Garnett, LeBron James, Tim
Duncan, Elton Brand, Luol Deng)
You can be an offensive star but nothing
special on defense (e.g., Gilbert Arenas,
Chauncey Billups, Steve Nash, Dirk
Nowitzki)
You can be a defensive star but nothing
special on offense (e.g., Jermaine ONeal or
Antonio McDyess)

Thank You

Вам также может понравиться