Crowdscouting/Individual Player Elo Ratings

[Buttonwood]

I just wanted to share a beta project using 'crowdscouting' and elo ratings.

The premise is people pick between 2 players and the project uses the elo formula to aggregate their opinion/'statistical fact' in a mathematically elegant way creating a time-series of each players 'rating.' If enough people use it, it would hopefully bake in many different factors: recent performance, statistical performance (fancy and otherwise), team circumstance, eye-ball test, etc.

In a few algorithms I should be about to reward users that pick guys ahead of the curve by increasing their influence. (ie. you championed McDavid before he was the best in the league) - but I think the fun will be in seeing how dynamic/useful the output data. It would produce career arcs, show player level consistency, etc..

 

[Doctor No]

I find this interesting as a concept.

How do you correct for bias in the voters (such as, say, a bunch of Avalanche fans flooding the site with picks for Duchene)? I've seen some odd results in the hockey-reference version of this.

 

[Doctor No]

Ah, I see - you've also got data on who the pickers were (unlike the H-R project).

 

[Buttonwood]

Right - hopefully in the first iterative people with police themselves.. their profile page returns their 'favored' teams 1-30, 30 players they 'champion, and 30 they 'bash,' so there is sort of a mirror there.

If they are hyping quite a few guys that turn out to be average, or their favorite team index completely out of line with the rest of the crowd/reality, then I will have the data to lessen their influence, but I haven't put that together yet. It would just be a formula that creates a constant 'k-factor' for each user. The smaller the 'k' the smaller weighting their influence on rankings. Since nobody has broken my trust, everyone is at k=1 right now.

 

[Teemu]

Hehe... I was just asked Bollig vs. Seguin

Now I just got Reto Berra v. Reto Berra.... I think I'm gonna go with "too close to call"

 

[Buttonwood]

Hehe... I was just asked Bollig vs. Seguin

Now I just got Reto Berra v. Reto Berra.... I think I'm gonna go with "too close to call"

Well thanks for working through some of the tough ones!

On the to-do list is implementing an algorithm to match similar players against each other. Position, team, age, and (maybe) current elo score will all be factors with some randomness thrown in.

I've also noticed same players matching up against each other, which goes against what I tell my php function to do (my similarity algo clearly isn't working either). Theres some black magic happening there, hopefully I can figure it out. Each way, thanks for contributing! I have a few other improvements as well (namely a player compare tool), and if anybody has any suggestions, let me know.

 

[Doctor No]

Now I just got Reto Berra v. Reto Berra.... I think I'm gonna go with "too close to call"

Was it "actual Reto Berra" vs. "whatever the heck the Avs saw in Reto Berra when they traded for him"?

 

[Buttonwood]

Was it "actual Reto Berra" vs. "whatever the heck the Avs saw in Reto Berra when they traded for him"?

Watching Berra in CGY, I saw MANY different versions of him.

COL definitely saw the 'Francois Allaire student Reto Berra' before that trade.

Also note if anybody is using the site - I need to update team rosters after some action yesterday. I've been doing this 'manually' with an excel VBA file then dumping it on the MYSQL server, does anybody know of some easy to read python or R code that might be able to automate this?

 

[Buttonwood]

Hehe... I was just asked Bollig vs. Seguin

Now I just got Reto Berra v. Reto Berra.... I think I'm gonna go with "too close to call"

Quick update: added a nice player similarity algorithm. Players should loosely match up based on position, team, and age. Also note their current elo rating has no current effect on the match-up so there will be a few lay-ups along the way...

 

[Steerpike]

The problem with this is it's just a complicated popularity contest. What's usually cool about numbers is they provide us with a view of players that is wholly impartial to their popularity.

 

[Doctor No]

The problem with this is it's just a complicated popularity contest. What's usually cool about numbers is they provide us with a view of players that is wholly impartial to their popularity.

Crowdsourcing is a common technique in predictive analytics. Check out Surowiecki's "The Wisdom of Crowds" for a fun introduction to the topic.

The trick here is in removing the bias, and aggregating those who are truly predictive.

 

[Buttonwood]

The problem with this is it's just a complicated popularity contest. What's usually cool about numbers is they provide us with a view of players that is wholly impartial to their popularity.

Like Doctor No mentioned, the idea is that a diversity of opinions leads to stronger predictive power. Following the principle: ensemble models (multiple models) perform better than a single models.. in a perfect world the community would be made up of fancy stat-heads and the 'watch the games' types OR everyone would have form their opinions based on the multitude of statistics.

The problem with 'dynamic' sports like hockey and football is that analytics have produced some great insights, but its fragmented and still needs context. We have all sort of things: good-old fashioned goal/point production, corsi for %, zone starts, zone entries, quality of teammates, quality of competition - all making up a big picture of a players value. The elo formula helps harness these statistical tidbit, if they can be put in a greater context.

Certainly, the popularity contest/unmitigated trolls worries me, but I have an algorithm to increase the influence of users that 'champion' risers and 'bash' fallers. Each users has a list of teams and players they favor, so I can work off of that (see below).

At the very least, it helps answer - is X overrated or underrated - well how is he rated in the first place? Advanced analytics are only super useful when there is a market (or popular opinion) inefficiency to take advantage of.

View attachment 84867

 

[Ishdul]

HF has the historical ones: http://www.hockey-reference.com/friv/ratings.cgi

It falls apart after the first bunch, and I remember not too long ago the results were thoroughly bizarre.

 

[Buttonwood]

HF has the historical ones: http://www.hockey-reference.com/friv/ratings.cgi

It falls apart after the first bunch, and I remember not too long ago the results were thoroughly bizarre.

Yes! I stumbled across those when researching this idea, but historical match-ups are prone to recency bias and the fact the NHL is a very different league decade to decade.

I think the elo application is better in a contemporary setting, a more ceteris paribus world.

Tennis is perfect for the application of elo, with actual head-to-head match ups. Hoping to create the same thing in hockey with the input being human judgement aided by statistical analysis.