Proper weighting of placings
Posted: Mon Nov 13, 2006 1:32 am
It is a hard thing- to rank people properly.
Too bad we don't know anyone who could do statistical analysis of the racers. Figure for each event-who statistically is likely to beat another based on past performance- perhaps weighting current races more .
In end you should get something that for sigma shows that x racer will beat y racer 68% of the time. (it may be less or more)
So from there you enter the races and a program figures out how to score each placing for points.
It sounds like a backwards way to work it out-
But I think it could work.
I think it would be an excellent thing for a Graduate statistics class to take on as a class project.
For instance lets say historically for GS we find Fluitt beats Pirnack 70% of the time. then Fluitt would rank above Pirnack- and lets say Pirnack almost always beats Vlad in GS then an order starts to formualte based on contest info- there have been so many contests that we could likely make some sense of the data.
Then the program figures out- how the placings of a contest should be weighted to arrive at the same ordering as statistics derived.
Then we use that model for the next 2 years for weighting contests until we have more data to process to refine it.
Eventually - we should get a pretty good model.
If things seem really out of a statistical range in the model...then we toss out that pair(s).
Just an idea...but soemone has a way to do this- and it probably evolved on paper by hand this way- but by compter we could get a very accurate way to do this and it would give us the flexability to change the model more easily as the numbers of participants increase.
[/i]
Too bad we don't know anyone who could do statistical analysis of the racers. Figure for each event-who statistically is likely to beat another based on past performance- perhaps weighting current races more .
In end you should get something that for sigma shows that x racer will beat y racer 68% of the time. (it may be less or more)
So from there you enter the races and a program figures out how to score each placing for points.
It sounds like a backwards way to work it out-
But I think it could work.
I think it would be an excellent thing for a Graduate statistics class to take on as a class project.
For instance lets say historically for GS we find Fluitt beats Pirnack 70% of the time. then Fluitt would rank above Pirnack- and lets say Pirnack almost always beats Vlad in GS then an order starts to formualte based on contest info- there have been so many contests that we could likely make some sense of the data.
Then the program figures out- how the placings of a contest should be weighted to arrive at the same ordering as statistics derived.
Then we use that model for the next 2 years for weighting contests until we have more data to process to refine it.
Eventually - we should get a pretty good model.
If things seem really out of a statistical range in the model...then we toss out that pair(s).
Just an idea...but soemone has a way to do this- and it probably evolved on paper by hand this way- but by compter we could get a very accurate way to do this and it would give us the flexability to change the model more easily as the numbers of participants increase.
[/i]