Weighted Stats?
- Thread starter skinneh
- Start date
Not sure I get you there. If you weight your stats over the last two weeks you'll find a lot of things, but certainly not how consistent is a player. You'll always end up with a few streaky players in your top 5 or 10. Or well, you could look at it in reverse and say that if a player is always in the top 20, he's consistent.
If you want to have only the trend from the last X games, it's pretty straightforward. Just give these games a decreasing weight. For example, if you want the last 10 games, just give a weight of 100% to the last game, 90% to the game before, 80% to the game before etc. When you get to 0%, the games just don't matter anymore.
However, if you want to keep the data for the whole season, then it's a different beast, because this data is cumulative. For example, if you leave a minimal weight of 10% for all the games, you'll slowly build a baseline over the season. So it will trend up and down a lot at the beginning, but not much at the end, because the 72 first games x 10% will outweight the last 10 games.
You could avoid that by saying that the last 10 games always count for, say, 50% of the score, but that would lead to other problems, especially in the beginning.
If you want to have only the trend from the last X games, it's pretty straightforward. Just give these games a decreasing weight. For example, if you want the last 10 games, just give a weight of 100% to the last game, 90% to the game before, 80% to the game before etc. When you get to 0%, the games just don't matter anymore.
However, if you want to keep the data for the whole season, then it's a different beast, because this data is cumulative. For example, if you leave a minimal weight of 10% for all the games, you'll slowly build a baseline over the season. So it will trend up and down a lot at the beginning, but not much at the end, because the 72 first games x 10% will outweight the last 10 games.
You could avoid that by saying that the last 10 games always count for, say, 50% of the score, but that would lead to other problems, especially in the beginning.
Thanks for the response. What your saying makes sense. I am looking more for players that are "hot" over the last few weeks.
I am writing a program to help me in my hockey pools
It will also look at team data and look for players that are hot, and have good matchups for the week.
It's a fun little side project
I am writing a program to help me in my hockey pools
It will also look at team data and look for players that are hot, and have good matchups for the week. It's a fun little side project
OK, I see what you want to do. Interesting.
One way of doing it would be to compute the cumulative PPG pace of a player and put it in a graph. Take the derivative of that graph for the last 10 games and voilà . Hot players will have a positive slope and guys in a slump will have a negative slope. And good players will always be higher on the graph.
So basically, your criterion here would be the current PPG of a player, minus his PPG 10 games ago. Something like that.
You could refine and calculate that slope more locally, see if it's noisy or what etc.
In a sense, it could be interesting to plot two graphs over #games played:
1-Cumulative PPG
2-First derivative (slope) of #1
Good players will be on top of #1, hot players will show spikes in #2.
One way of doing it would be to compute the cumulative PPG pace of a player and put it in a graph. Take the derivative of that graph for the last 10 games and voilà . Hot players will have a positive slope and guys in a slump will have a negative slope. And good players will always be higher on the graph.
So basically, your criterion here would be the current PPG of a player, minus his PPG 10 games ago. Something like that.
You could refine and calculate that slope more locally, see if it's noisy or what etc.
In a sense, it could be interesting to plot two graphs over #games played:
1-Cumulative PPG
2-First derivative (slope) of #1
Good players will be on top of #1, hot players will show spikes in #2.
wgknestrick
Registered User
- Aug 14, 2012
- 5,867
- 2,608
Be careful in trying to "weight" based on small sample size. You are basically amplifying the "noise" and variation IMO because your weight "groups" are too small. What would make a game in Nov be less valuable than a game in Feb? They are "equal" and should be assumed "equal" to be able to do any trend analysis.
Weighting is generally saved for trying to normalize stats where players were under different circumstances (ie different teams), not trying to spot trends. There are already established methods for that
http://en.wikipedia.org/wiki/Trend_estimation
Sounds like you're looking for a weighted moving average.
http://en.wikipedia.org/wiki/Moving_average#Weighted_moving_average
http://en.wikipedia.org/wiki/Moving_average#Weighted_moving_average
Thanks for the link! I am reading through it now. In my opinion the points in Nov are less valuable then the points in Feb for fantasy sports at least (Yahoo pool, Fanduel).
Example (changed the months):
1. Grabovski had 8 points in Oct. He has 0 points through the end of Nov. Total points for 2 months = 8. (maybe he is hurt, maybe he has less PP time, maybe he was moved to the third line, maybe his wingers got injured) for whatever reason 0 points in Nov.
2. Tatar had 0 points in Oct. Nov comes around and he is moved to the second line, or is maybe getting some pp time. He has 8 points through the last 2 weeks of Nov. Total points for 2 months = 8;
I would want Tatar no? His trend is upwards. Grabovski's is downward. Tatar is hot for whatever reasons, Grabovski is cold for whatever reasons.
Right now I need to look at each player and their stats and try and compare them in my small brain. Maybe I am looking at all second liners in the NHL. I was hoping to use some sort of weighted stat that would compare x number of players and give me an idea of which players have been better lately.
Maybe Weighted isn't the way to go. Maybe I am way off base.
Thanks again for your guys input.
wgknestrick
Registered User
- Aug 14, 2012
- 5,867
- 2,608
Thanks for the link! I am reading through it now. In my opinion the points in Nov are less valuable then the points in Feb for fantasy sports at least (Yahoo pool, Fanduel).
Example (changed the months):
1. Grabovski had 8 points in Oct. He has 0 points through the end of Nov. Total points for 2 months = 8. (maybe he is hurt, maybe he has less PP time, maybe he was moved to the third line, maybe his wingers got injured) for whatever reason 0 points in Nov.
2. Tatar had 0 points in Oct. Nov comes around and he is moved to the second line, or is maybe getting some pp time. He has 8 points through the last 2 weeks of Nov. Total points for 2 months = 8;
I would want Tatar no? His trend is upwards. Grabovski's is downward. Tatar is hot for whatever reasons, Grabovski is cold for whatever reasons.
Right now I need to look at each player and their stats and try and compare them in my small brain. Maybe I am looking at all second liners in the NHL. I was hoping to use some sort of weighted stat that would compare x number of players and give me an idea of which players have been better lately.
Maybe Weighted isn't the way to go. Maybe I am way off base.
Thanks again for your guys input.
I think I am starting to see what you are trying to accomplish with this.
Basically you are trying to find depth players who (most likely) may have been placed in significantly more favorable circumstances with respect to:
-increased PP time (maybe they are filling in for injured player on top unit)
-Better linemates (maybe they are filling in for injured player on 1st/2nd line)
-Good health
-Favorable O-zone starts
vs the con for each above
You aren't really worried about the "factors" above that determine the change in productivity (it is a very complex system), but are just looking for the end result (much easier to estimate), as output. This would probably work in the long haul, but I would guess it would "lag" behind as you will be deciding based on circumstances that may not be true as players return to health and roles change back. Almost like deciding to fill your car up after you run out of gas, instead of looking at the gas gauge.
I don't really believe in "hot or cold streaks". Everything has a cause/effect whether it is just perceived by the observer and doesn't really exists, if there is an underlying injury reason, or if it is just natural player development relative to these curves below.
http://hockeyanalytics.com/Research...boards.mandatory.com/showthread.php?t=1238041
I think all players have a "base skill level vs age" and significant variations from that are generally well explained in hindsight. Too bad most injuries aren't even reported now to help us.
This is exactly what I am trying to accomplish. I will still have to do secondary research on the players. Why is their production higher? why are they getting more points? Will their bonus production drop because x player has come back from injury? But if I can narrow my search from 50 players to 5 for that week, or that day, then it's a lot less work.
At least that is my first idea on what I should try accomplish with this application. Once I get the data and see it, on a day to day basis, or a week to week basis, it will be interesting to see if it is helping me or not
. I may then need to reevaluate my methods!
