Burkie's Challenge Accepted: Lets Make An Analytics System!

[trublmaker]

Originally Posted by AcGold

2 degrees. Plenty of stats courses and knowledge. Start by understanding the problem. Find correlations in data between one desired factor and another. Burke wants to know about drafting, come up with stats on certain variables. Find the correlation of players drafted in the past with rates of success, points produced, time played etc. to predict the success probability of future prospects based on their available stats. Advanced stats basically means find correlations in calculations with more than 2 variables.

show off

[kirant]

Originally Posted by AcGold

2 degrees. Plenty of stats courses and knowledge. Start by understanding the problem. Find correlations in data between one desired factor and another. Burke wants to know about drafting, come up with stats on certain variables. Find the correlation of players drafted in the past with rates of success, points produced, time played etc. to predict the success probability of future prospects based on their available stats. Advanced stats basically means find correlations in calculations with more than 2 variables.

I think this comes down to two key aspects:
- What we want to call a successful player. Is it points? Relative Corsi or other advanced stats? Maybe something with a much more nuanced calculation?
- How we can identify a successful player from prospect level scouting. This might be the harder of the two points for a hockey forum IMO since, unless someone has a validation method or lots of old tapes, we can't prove the capability of the system generated. The data of drafted players, such as those from the CHL, is pretty limited as they don't keep records of quite a few stats. For example, I think the WHL only tracks points, PIM, and +/-.
http://www.whl.ca/roster/gameByGame/id/10490

That being said, I'm willing to help out wherever possible. Just wrapped up an Engineering degree so I'm mot sure there's much practical application I can provide...but again, would be interested in helping out.

[AcGold]

don't look for one correlation, look for many. In a statistical analysis such as this don't look for one indicator or a threshold but a collection of correlations to see which statistics hold the greatest correlation. it wouldn't be hard, just take the available information and build a matrix multiplier that can tell which relationships have the greatest correlation coefficients.

what do we want to call a successful player? don't look at it like that, that's not the smartest way to go about it. It would be better to find the strongest correlations.

[trackercowe]

Originally Posted by photon

We should start with a statistician. I don't even know one let alone one that posts on CP.

So we need these guys?



There's a mathematician, a different kind of mathematician, and a statistician.

[Derek Sutton]

Moneyball it. Draft skaters with the highest shooting percentages and goalies with the highest save percentages.

[Wastedyouth]

Originally Posted by Iceman90

Brian Burke promised to pay cash for an advanced analytic system that will help them draft better and/or make the team better. I'd like nothing more than to take Brian Burke's cold hard cash. I figure lets crowd source this super awesome analytics system, since we have some of the brightest minds in hockey here on CP.

So, who has an idea of where we should start?

I LOL'd

[Alberta_Beef]

It's impossible to make an analytics system that will be up to the level needed. This isn't baseball where almost everything can be quantified and calculated with stats, there are simply too many parts of the game where opinions would factor in.

Hits, Giveaways, Takeaways, Quality of Competition... etc are all based on human opinion.

[Wastedyouth]

Originally Posted by Derek Sutton

Moneyball it. Draft skaters with the highest shooting percentages and goalies with the highest save percentages.

I think you missed the point of moneyball.

[Alberta_Beef]

Originally Posted by Wastedyouth

I think you missed the point of moneyball.

I think you missed the joke

[Wastedyouth]

Originally Posted by Alberta_Beef

It's impossible to make an analytics system that will be up to the level needed. This isn't baseball where almost everything can be quantified and calculated with stats, there are simply too many parts of the game where opinions would factor in.

Hits, Giveaways, Takeaways, Quality of Competition... etc are all based on human opinion.

None of the bolder are utilized anywhere in advanced stats because they are indicators of nothing and are also poorly tracked.

QofC isn't based off of human opinion at all. It's based off of the Corsi of the opposing players on the ice.

[RisebroughRuinedMyYouth]

Originally Posted by TheDebaser

I've thought about making a stat that could help figure out a shots quality, which could be used to figure out which players are getting good shots off rather than just measuring the amount of pucks hitting the goalie. I could see it being done with just an x,y position from the goal.

There are two ways it could be done.

1)An equation that uses the xy position of a shot to get the angle/distance of a shot to assign it a score. Something like "Quality= Sin(Angle)/Distance" where "Angle = InvTan(y/x)" and "Distance = x^2 + y^2" would output a rough score. More work would be needed.

2)Use tons of data from the entire league to create a kind of heatmap that could assign a shooting % from a given part of the ice. This would be ideal, but it would probably require programming knowledge. (If the data is even out there.)

These stats could be used to measure a player offensively and defensively. Instead of counting shots for and against, count shooting % opportunities for and against.

Ex: If a player gets a 9% and a 15% opportunity in a shift, but the opponent gets a 12% and 14% opportunity, his score would be -.02 for his shift. (I think there's room to improve this with stats knowledge.)

This system would show if a player is not getting tons of chances, but is getting better ones, or if he is keeping players to the outside defensively. Since it's the same idea as corsi, but with shot quality factored in, this is also a possession stat.

What do you guys think?

Would love to see a Kirk Goldsberry-type map for hockey (he does basketball shooting heatmaps). I wonder if teams possess proprietary versions.

[Alberta_Beef]

Originally Posted by Wastedyouth

None of the bolder are utilized anywhere in advanced stats because they are indicators of nothing and are also poorly tracked.

QofC isn't based off of human opinion at all. It's based off of the Corsi of the opposing players on the ice.

If you are going to use it for drafting a player those things need to be considered and quantified. Which they cannot be. Which is exactly why no analytics system will be good enough

[SebC]

Originally Posted by AcGold

Advanced stats basically means find correlations in calculations with more than 2 variables.

In the context of this discussion, advanced stats basically means anything that doesn't show up on a boxscore.

[CalgaryFan1988]

Originally Posted by Wood

I'll supervise

And I'll make sure everyone does their job safely.

Who is in charge of payroll?

[DeluxeMoustache]

Originally Posted by TheDebaser

I've thought about making a stat that could help figure out a shots quality, which could be used to figure out which players are getting good shots off rather than just measuring the amount of pucks hitting the goalie. I could see it being done with just an x,y position from the goal.

There are two ways it could be done.

1)An equation that uses the xy position of a shot to get the angle/distance of a shot to assign it a score. Something like "Quality= Sin(Angle)/Distance" where "Angle = InvTan(y/x)" and "Distance = x^2 + y^2" would output a rough score. More work would be needed.

2)Use tons of data from the entire league to create a kind of heatmap that could assign a shooting % from a given part of the ice. This would be ideal, but it would probably require programming knowledge. (If the data is even out there.)

These stats could be used to measure a player offensively and defensively. Instead of counting shots for and against, count shooting % opportunities for and against.

Ex: If a player gets a 9% and a 15% opportunity in a shift, but the opponent gets a 12% and 14% opportunity, his score would be -.02 for his shift. (I think there's room to improve this with stats knowledge.)

This system would show if a player is not getting tons of chances, but is getting better ones, or if he is keeping players to the outside defensively. Since it's the same idea as corsi, but with shot quality factored in, this is also a possession stat.

What do you guys think?

I like the intent.

Shot location is unfortunately inadequate as a proxy for shot quality. The cross crease tap-in right beside the post may have the same xy as the puck being jammed into a firmly planted goalie's pad. One has a 90 plus percent chance of going in, the other closer to zero.

In that case, or for many goals in close, there are many (as far as I know) unmeasured items such as, I don't know... location of prior puck possession, goalie position, and whatever the heck happened leading up to that shot.

Then there are situations where a point shot is made with a lot of net front traffic compared to one where the goalie has a clear line of sight.

You may almost have to get to the point of scoring player positions for all teams to identify relative team position quality (?) which may identify lapses in coverage.

You have to distinguish between manpower situations, and then beyond that,
Corsi I believe went to the point of considering score effects (is a team up by 2 or trailing by 1 and pressing). This stuff gets incredibly complicated, especially when the players' mantra is to get pucks on net and get traffic

The one that seems easiest to identify is if that if the shot is from the neutral zone, identifying if the shooter is or is not McGrattan.

[Wastedyouth]

Originally Posted by Alberta_Beef

If you are going to use it for drafting a player those things need to be considered and quantified. Which they cannot be. Which is exactly why no analytics system will be good enough

Good enough for what?

No analytics system will ever be perfect. Its a tool. If there was a system to accurately predict everything, then whats the point of playing the game?

[Where ru Chris O'Sullivan]

Originally Posted by Alberta_Beef

It's impossible to make an analytics system that will be up to the level needed. This isn't baseball where almost everything can be quantified and calculated with stats, there are simply too many parts of the game where opinions would factor in.

Hits, Giveaways, Takeaways, Quality of Competition... etc are all based on human opinion.

And in Baseball there is no physical interaction with other players. This is a huge part of the NHL. It's not only being able to skate, pass, shoot, but the ability to workaround others (the best in the world).

[EldrickOnIce]

Burke's answer is perfect regarding 'advanced' stats: Good for support, not necessarily illuminating.
Crosby has been almost unanimously regarded as the best player in the world for some years now. So while there will be anomalies due to sample size, etc - in the long run that all should be normalized and the analysis good.
And so here's the analysis:
Crosby's long term advanced stats are, by and large, not statistically different than Michael Backlund's.
A system that is supposed to identify/predict the 'best player' should do just that - and corsi, etc, obviously does not. It identifies something - and something valuable in terms of outshooting/chancing - so as some small part of an overall evaluation package, it is excellent.
And imo should be used as Treliving said - we will look at everything, all the information, including analytics

[Maritime Q-Scout]

Are we just talking the draft here?

If so, here's my proposal.

Run an analysis on all NHL players from the last 15 years to see where they played their draft year. Compare that to who their coaches were. Then how successful they were in the NHL.

I'm free balling here.

Variables to matrix

NHL draft ranking by central scouting
Junior team
Junior coach
Games played
Playoff games played
Winning percentage
Playoff winning percentage
Points per game (for offence)
Blocked shots (for defence)
Defensive zone giveaways
Total ice time
Time with the puck
Completed pass percentage
Body checks per game
Penalty minutes per game

Now I'd have to come up with an exact forumla.

Any other key variables that should be considered?

[BACKCHECK!!!]

I can't wait to see all the new revolutionary ways of telling which players have had a good season.