2016-17 Team Win Projection Contest / Discussion

[ampersand5]

Thanks for the response Nate, and I apolagize if I wasn't sufficiently clear.

I'll try and rephrase my question in a simpler way.

My model says team A has a 33% chance of beating team B.

However, my model is only 75% accurate in its forecasts.

Lets say that Joe Smith comes up to me and says "Hey Daniel, how confident are you that team A beats team B tomorrow.

Do I say 33% because that's what my model says? or can I incorporate the uncertainty of my model into my confidence of the game's outcome?

[Nate]

...
However, my model is only 75% accurate in its forecasts
...

I don't really understand what that means. Can you elaborate?

...We can look at the RMSE from previous year's predictions to get a sense at how accurate our model is...

The total number of wins in the league over the course of a season is fixed. So even if the RSME is large, the net average error will be 0.

Something you can do is to look for patterns in the error - for example check whether are the errors typically toward 41 wins or away from 41 wins - and then incorporate that knowledge into your predictions. We could, in some sense, say that a model which errs toward 41 is underconfident, and a model which errs away from 41 is overconfident.

[Mike G]

Using Kevin P's minutes projections found here --
viewtopic.php?f=2&t=9211
-- I've plugged in last year's RPM, BPM, and WinShares rates, along with my own eWins and faux RPM, to project wins for the season.
RPM returns exactly 1230 wins, 41*30; the others are all higher, so they're scaled down to avg 41.

Code: Select all

tm    BPM   RPM   fRPM   WS    eW
GSW    69    69    68    71    78
SAS    62    62    62    63    70
Cle    58    61    58    55    55
LAC    51    54    52    53    57
Tor    53    52    51    51    53

Okl    51    48    51    50    45
Bos    50    50    48    47    43
Uta    50    49    46    47    46
Por    46    45    43    44    43
Det    41    47    43    40    49

Atl    41    46    47    42    41
Hou    45    49    45    41    35
Chi    44    40    43    42    44
Min    42    39    44    40    46
Cha    42    40    42    44    41

tm    BPM   RPM   fRPM   WS    eW
Ind    37    39    42    38    51
Mem    41    38    37    40    41
Dal    42    39    39    43    33
Was    36    40    39    35    44
Den    35    35    37    31    38

Orl    35    33    31    38    37
Sac    34    35    33    34    31
NOP    32    34    32    35    34
Mia    34    34    30    39    26
NYK    32    34    34    29    31

Mil    29    27    32    35    34
Phx    30    26    28    28    25
LAL    26    21    28    27    22
Brk    21    24    22    25    18
Phl    21    19    24    23    19

avg    41    41    41    41    41

These projections are not regressed toward the mean.

Avg departure from the 5-metric avg:

Code: Select all

BPM    RPM    fRPM    WS     eW
1.62   1.99   1.47   2.03   3.47

eWins has either the high or low projection for almost every team, un-regressed.
Players with less than 250 minutes or so didn't have an RPM that I could find; so I just used their BPM in the RPM slot.
http://www.basketball-reference.com/lea ... anced.html

Rookie rates are guessed at by their rank in their teams' minutes distribution: Players with more minutes had generally better numbers, and they are supposed to be better than those with fewer minutes.
Players who missed last season are treated the same as rookies. Player age is not accounted for, and minutes may be adjusted later.

[Crow]

That is useful info. Thanks for the work.

[shadow]

Here are my projections, using 3 years of RPM (weighted 6-3-1), a simple aging curve and KP's minutes.

EDIT: I've changed this to use whole wins, per Crow's recommendation. Also fixed rounding issue so wins total to 1230.

Code: Select all

GSW	66
CLE	59
SAS	57
LAC	50
BOS	49
UTA	49
HOU	49
TOR	48
OKC	46
POR	44
MEM	42
WAS	42
DET	41
ATL	41
CHA	41
DAL	41
SAC	40
CHI	39
IND	39
MIA	38
NOP	37
ORL	37
DEN	36
NYK	36
MIN	33
MIL	29
PHX	28
BRK	28
LAL	24
PHI	21

[Crow]

Likeable methodology.

I will say in general that I don't think fractions of a win belong in the contest. Vegas has a market making reason to straddle but here it probably gives some advantage vs. projections that use whole numbers.

[Crow]

My suggestion may or may get universal implementation.

Shadow, it appears you are missing 3 wins leaguewide. Perhaps from the rounding.

[BasketDork]

My 2016-2017 predictions...

EASTERN CONFERENCE
Cleveland..........57
Toronto............ 52
Boston...............50
Atlanta...............45
Charlotte............44
Indiana...............43
Washington....... 42
Chicago..............41
Detroit................40
Miami.................38
New York............35
Orlando...............33
Milwaukee..........28
Brooklyn.............20
Philadelphia.......19

WESTERN CONFERENCE
Golden State........67
San Antonio........ 59
L.A. Clippers........53
Houston...............51
Oklahoma City.... 48
Utah......................47
Portland............... 44
Memphis...............43
Dallas.....................40
Minnesota.............38
Sacramento...........36
Denver....................35
New Orleans..........32
Phoniex..................29
L.A. Lakers.............21

When is the cutoff for amendments ?

EASTERN CONF. PLAYOFFS :
(1) Cleveland vs. (8) Chicago
(4) Atlanta vs. (5) Charlotte
(3) Boston vs. (6) Indiana
(2) Toronto vs. (7) Washington

WESTERN CONF. PLAYOFFS
(1) Golden State vs. (8) Memphis
(4) Houston vs. (5) Oklahoma City
(3) L.A. Clippers vs. (6) Utah
(2) San Antonio vs. (7) Portland

[Crow]

Tip-0ff of first game unless you have a good non-advantage seeking excuse and are quick to use it.

[Crow]

Team
Atlanta Hawks* 46
Boston Celtics* 52
Brooklyn Nets 23
Charlotte Hornets* 39
Chicago Bulls 42
Cleveland Cavaliers* 57
Dallas Mavericks* 37
Denver Nuggets 36
Detroit Pistons* 39
Golden State Warriors* 70
Houston Rockets* 47
Indiana Pacers* 39
Los Angeles Clippers* 55
Los Angeles Lakers 23
Memphis Grizzlies* 44
Miami Heat* 35
Milwaukee Bucks 29
Minnesota Timberwolves 37
New Orleans Pelicans 32
New York Knicks 37
Oklahoma City Thunder* 47
Orlando Magic 31
Philadelphia 76ers 19
Phoenix Suns 37
Portland Trail Blazers* 42
Sacramento Kings 34
San Antonio Spurs* 56
Toronto Raptors* 48
Utah Jazz 48
Washington Wizards 46

Total 1230

Might tweak it in next 48 hours.

Really I want to give out about 100 less wins but I have to give them to someone.

[nrestifo]

Here are mine:

I retweeked them a little today from the projections posted in my recent Nylon Calc article to account for the Snell/MCW trade and to give less weight to older games in my my modeling approach.

Almost no difference, which makes total sense to me.

My projections for the 2016-2017 NBA season below.

win_team round_wins
GSW 70
CLE 61
SAS 60
LAC 52
OKC 52
TOR 50
BOS 48
UTA 47
HOU 44
WAS 43
CHA 43
MEM 43
CHI 42
DAL 42
POR 42
ATL 41
IND 39
NOP 39
SAC 39
NYK 38
DET 37
MIN 32
ORL 32
DEN 31
MIA 31
MIL 30
PHX 29
BKN 26
PHI 24
LAL 23

Total 1230

Best of luck everyone.

[mtamada]

...
However, my model is only 75% accurate in its forecasts
...

I don't really understand what that means. Can you elaborate?

There are a number of ways to measure the uncertainty of a forecast. Bottom line for ampersand5 is simply this: regression to the mean.

Meaning that the more uncertain (or less accurate) your forecast, the more you should adjust your forecast from the 33% in your example to 50%. In the extreme case where the model provide no information at all, all that you can say is that Team A has a 50% chance of beating Team B, and that's your forecast for every single game.

Whereas if you knew for certain that A has a 33% chance of beating B, then your forecast should be 33%.

For situations that are in between (i.e. all practical real-world situations), regress a certain proportion of the distance between the two.

There's a simple formula for figuring out how much to regress when we're dealing with normally distributed random variables (Tom Tango cites it often in his blog); I don't know what the analogous formula is for probabilities. It might be equally simple or it might not be, due to the multiple different ways of measuring the uncertainty of a probabilistic forecast.

[sbs]

GSW 69.1
CLE 57.2
SAS 55.5
LAC 51.0
BOS 50.6
TOR 50.3
UTA 48.4
OKC 45.4
POR 45.3
HOU 45.3
DET 42.8
CHA 42.3
IND 41.2
WAS 41.1
ATL 40.9
CHI 39.8
MIN 39.6
MEM 39.1
DAL 37.7
ORL 36.8
DEN 36.5
SAC 35.4
NOP 35.4
NYK 35.1
MIA 34.6
MIL 33.0
PHX 28.8
LAL 24.3
BKN 23.9
PHI 23.8

[Crow]

Another opinion, if you don't post your predictions here you can"t "win the contest". You can be better, you might be tracked, but if you don't post it here you didn't "enter" this contest.

[GK5]

First entry after watching for a few years. Best of luck to all.

GSW 73
SAS 53
CLE 52
TOR 52
UTA 51
LAC 49
BOS 48
OKC 48
HOU 45
POR 45
CHA 43
CHI 41
DET 40
MIN 40
DEN 39
IND 39
WAS 39
ATL 38
DAL 38
SAC 38
MEM 37
ORL 37
MIA 36
NOP 35
NYK 31
PHO 31
MIL 30
PHI 29
BRK 28
LAL 25

Total 1230