paper: Weeks 1-2 Elo Analysis

[Ether]

Originally Posted by Michael Hill

The file should be a tab delimited file of red1 red2 red3 blue1 blue2 blue3 redscore bluescore

Would you please post a ZIP of your input data file?

[Michael Hill]

Originally Posted by Ether

Would you please post a ZIP of your input data file?

It's the same one I PMed you a couple weeks ago, but sure.

https://dl.dropboxusercontent.com/u/5193107/scores.zip

You'll need to install the trueskill package ('pip install trueskill') to use it.

[Michael Hill]

So I decided to take this data a little bit further. What I did was to take all of these calculations and run them through the data this year and try to predict matches. For this, I also included a modified Elo system that has diminishing returns for large margins of victory (Calling this Elo Mod). I got some rather surprising results.

My baseline was just using OPR for predicting match outcomes, it was able to predict about 77.1% of the matches this year. This was calculated by adding up the OPRs of each alliance and comparing with the result of the match. TrueSkill was able to predict 79.0% of the matches, a pretty good improvement. I need to develop the prediction model a bit better because it currently doesn't take into account the standard deviation as a measure of certainty. The modified Elo system was able to predict 79.5% of matches, an improvement over TrueSkill. The baseline, unadulterated Elo system as used in this thread was able to predict a whopping 81.4% of matches, by far the best out of any of these models. There is still room for improvement with the TrueSkill and Modified Elo. With the modified Elo, there are some constants that can be tuned for better results. But overall, the results are somewhat interesting. It seems that no matter the ranking model used, about 1 in 5 qualification matches will result in an upset.

Here is the spreadsheet I used: https://dl.dropboxusercontent.com/u/...Trueskill.xlsx

[Ether]

Originally Posted by Michael Hill

What I did was to take all of these calculations and run them through the data this year and try to predict matches.

Since you used the word "predict", is it safe to say that what you did was to use the data from weeks 1 through n-1 to predict the matches in week n?

[Michael Hill]

Originally Posted by Ether

Since you used the word "predict", is it safe to say that what you did was to use the data from weeks 1 through n-1 to predict the matches in week n?

No, I guess I should have used a better word than "predict". More like "postdict". I went into it "knowing" an Elo/TrueSkill rating and tested against the data that was used to calculate it. Of course we won't have all the data to calculate the final Elo ratings during the season. My next step is to calculate all the ratings as if it is just before championships and then see how each does with "predicting" championship matches. My SWAG is that the success rate in predictions using the "postdicted" matches is a ceiling for how good we can hope for the predictions to be.

[Ether]

Originally Posted by Michael Hill

My baseline was just using OPR for predicting match outcomes, it was able to predict about 77.1% of the matches this year. This was calculated by adding up the OPRs of each alliance and comparing with the result of the match. TrueSkill was able to predict 79.0% of the matches, a pretty good improvement. I need to develop the prediction model a bit better because it currently doesn't take into account the standard deviation as a measure of certainty. The modified Elo system was able to predict 79.5% of matches, an improvement over TrueSkill. The baseline, unadulterated Elo system as used in this thread was able to predict a whopping 81.4% of matches, by far the best out of any of these models.

Try calculating "OPR" using min L1 norm of residuals (LAD) instead of min L2 norm (least squares), and see how that compares.

[Michael Hill]

I've tried the L1 optimization problem, but l1-magic is giving me fits. For some reason, it blows up after 10 iterations.

However, I have done the analysis that I wanted. I calculated all the rating systems using data from events prior to CMP, then I used that data to predict CMP matches.

In summary:
OPR: 72.46% Correct
TrueSkill: 69.31% Correct
Elo: 72.90% Correct
Elo Mod: 71.71% Correct

TL;DR: We're okay, but not great at predicting matches. OPR is okay at it, but Elo is better.

I'm still somewhat surprised that Elo is slightly better.

Updated Spreadsheet: https://dl.dropboxusercontent.com/u/...skill%202.xlsx

[Ether]

Originally Posted by Michael Hill

I've tried the L1 optimization problem, but l1-magic is giving me fits. For some reason, it blows up after 10 iterations.

What solver are you using ?

[Ether]

Originally Posted by Michael Hill

In summary:
OPR: 72.46% Correct
TrueSkill: 69.31% Correct
Elo: 72.90% Correct
Elo Mod: 71.71% Correct

Try CCWM.

[Michael Hill]

Originally Posted by Ether

What solver are you using ?

I'm using the l1eq_pd.m function in the L1-Magic library (http://users.ece.gatech.edu/~justin/l1magic/)

Originally Posted by Ether

Try CCWM.

CCWM: 71.41%

[Ether]

Originally Posted by Michael Hill

I'm using the l1eq_pd.m function

That's the wrong solver.

% l1eq_pd.m
%
% Solve
% min_x ||x||_1 s.t.  Ax = b 

Firstly, you cannot find a min L1 norm vector x such that Ax=b because there is no vector x such that Ax=b, since the system is overdetermined.

Secondly, what you want to find is the min L1 norm of the residuals, not of the solution vector itself.

For the set of overdetermined linear equations Ax ≈ b, x is the solution vector. The residuals are b-Ax. So you want find a solution vector x which minimizes the L1 norm of b-Ax.

[Ether]

Attached is a comparison of b-Ax residuals for L2 and L1 OPR.

Alliance scores computed from L1 OPR are within +/-10 points of the actual scores 33.5% of the time.

Alliance scores computed from L2 OPR are within +/-10 points of the actual scores only 22.4% of the time.

It is on that basis that I postulate that L1 OPR might be a better predictor of match outcome.

[EDIT]Cannot add attachments to threads associated with papers. Brandon: can you please change this setting to allow attachments? Thank you.[/EDIT]

[Ether]

Originally Posted by Michael Hill

My baseline was just using OPR for predicting match outcomes, it was able to predict about 77.1% of the matches this year.

I just ran the OPR numbers using the data you linked in your earlier post. I came up with 6919 of 8921 matches correctly "postdicted", or 77.56%

Our numbers are very close, but I had expected them to be identical.

Here's a link to an XLS spreadsheet.

[Michael Hill]

Originally Posted by Ether

I just ran the OPR numbers using the data you linked in your earlier post. I came up with 6919 of 8921 matches correctly "postdicted", or 77.56%

Our numbers are very close, but I had expected them to be identical.

Here's a link to an XLS spreadsheet.

The differences are how we're calling tied matches. The way you're calculating if OPR "correctly" predicted falls apart when the actual result of the match was a tie. Almost always, the OPR will not predict a tie.

[Michael Hill]

Originally Posted by Ether

Attached is a comparison of b-Ax residuals for L2 and L1 OPR.

Alliance scores computed from L1 OPR are within +/-10 points of the actual scores 33.5% of the time.

Alliance scores computed from L2 OPR are within +/-10 points of the actual scores only 22.4% of the time.

It is on that basis that I postulate that L1 OPR might be a better predictor of match outcome.

[EDIT]Cannot add attachments to threads associated with papers. Brandon: can you please change this setting to allow attachments? Thank you.[/EDIT]

Ya, I'll admit I don't know what I'm doing with the L1 stuff. It's a new concept to me, so I just did some quick googling and thought I found a quick solution.