Calgarypuck Forums - The Unofficial Calgary Flames Fan Community - View Single Post - Burkie's Challenge Accepted: Lets Make An Analytics System!

cannon7 · 05-03-2014, 12:37 PM

I'm a Computer Scientist with a background in ML (machine learning). I've spent some time looking into the potential for leveraging ML to find patterns for exactly this purpose. Unfortunately hockey isn't as quantifiable as, say, baseball -- so the Moneyball approach is far more limited. But it is still possible, assuming you can get the data you need.

Here's the bird's eye view...

It all starts with an algorithm, or a mathematical equation, where the input is a player's pre-draft statistics, the league they are in, the team they play from, where they were born, when they were born, height/weight (overtime, preferrably) and so on. the output is an accurate probability for success in the NHL.

The way you build this algorithm for ML purposes is that you provide all the above inputs and their respective properties. For example, you wouldn't ust have Petergorough, ON as a property -- instead you'd have the number of players born there, the number of players who made the NHL who were from there, etc. You plug these variables in, provide some basic rules (logic) by which the ML can alter the algorithm (of which there should be millions/billions of combinations) and then run it.

Now the way ML will determine if it has improved the algorithm or not is to compare the input with the output. So it will crunch all the existing data you can get your hands on as to a player's pre-draft stats and post-draft NHL success. It will then compare the output of the algorithm with reality. Now it does this in aggregate, meaning it compares thousands of inputs with thousands of outputs, since comparing just one player to reality would only be anecdotal.

Eventualy (billions of calculations later) it will come to a point where it cannot improve the algorithm any more given the dataset provided. You then take this algorithm and apply it to the current draft crop and see what you get.

05-03-2014, 12:37 PM	#46
cannon7 Needs More Cowbell Join Date: Aug 2002 Location: Not Canada, Eh? Exp:	I'm a Computer Scientist with a background in ML (machine learning). I've spent some time looking into the potential for leveraging ML to find patterns for exactly this purpose. Unfortunately hockey isn't as quantifiable as, say, baseball -- so the Moneyball approach is far more limited. But it is still possible, assuming you can get the data you need. Here's the bird's eye view... It all starts with an algorithm, or a mathematical equation, where the input is a player's pre-draft statistics, the league they are in, the team they play from, where they were born, when they were born, height/weight (overtime, preferrably) and so on. the output is an accurate probability for success in the NHL. The way you build this algorithm for ML purposes is that you provide all the above inputs and their respective properties. For example, you wouldn't ust have Petergorough, ON as a property -- instead you'd have the number of players born there, the number of players who made the NHL who were from there, etc. You plug these variables in, provide some basic rules (logic) by which the ML can alter the algorithm (of which there should be millions/billions of combinations) and then run it. Now the way ML will determine if it has improved the algorithm or not is to compare the input with the output. So it will crunch all the existing data you can get your hands on as to a player's pre-draft stats and post-draft NHL success. It will then compare the output of the algorithm with reality. Now it does this in aggregate, meaning it compares thousands of inputs with thousands of outputs, since comparing just one player to reality would only be anecdotal. Eventualy (billions of calculations later) it will come to a point where it cannot improve the algorithm any more given the dataset provided. You then take this algorithm and apply it to the current draft crop and see what you get.