Biggest upsets of 2016

[Torvando]

Originally Posted by jajabinx124

Wow, I didn't expect that semi-final match from the 2016 10,000 lakes regional to be on there. I'm surprised the win likelihood was only 2.6% for blue alliance.

That is surprising to me as well. I thought our alliances were quite even with your alliance having a slight advantage. It would have been fun to see if we could have won if we didn't have technical problems.

[Caleb Sykes]

Originally Posted by jajabinx124

Wow, I didn't expect that semi-final match from the 2016 10,000 lakes regional to be on there. I'm surprised the win likelihood was only 2.6% for blue alliance.

SF 1-1 had the blue alliance with a 11.3% chance to win, but replacing 2502 with 2470 dropped the chances quite a bit for SF 1-2.

[blueyoshi256]

I'm actually not that surprised by the 10,000 lakes match.
I mean, on the red alliance, you had 2052 and 525, with a combined total of 1 loss (by 525 to 2052). Each of them was scoring like 5+ high goals average.
And on the blue side...
2502, who were capable of scoring 5+ high goals easy, but were broken
2823, who could score 5 high goals, but only if we were really lucky. A good match would be like 2, and 0 was pretty common at 10k
3038, who were an awesome breacher and 3rd bot, (with no high goal scoring)
and 2470, who also couldn't score a high goal (but turns out that they were great at defense!)

So, I'd estimate a 10 high goal difference on average between the alliances going into the match. That's an insane deficit.

[Whatever]

I think the other odd thing to note about that match was 2470 went into the elimination rounds as the 4th replacement available. The reason they were out there was because there had already been two robots swapped in. I haven't actually heard what happened to the 3rd team. I thought it really spoke highly of 2470 that they were ready for the match.

[weaversam8]

Hey Caleb!

If you're working with prediction models for FRC, I'd suggest you checkout what my team, The Wired Wizards, did with ORB.

The link to the forum post is here: https://www.chiefdelphi.com/forums/s...d.php?t=147817

We worked with match prediction specific to Stronghold, doing analysis based on teams affinity with different defenses... there's more information in the post...

Good luck!

[Caleb Sykes]

Originally Posted by weaversam8

Hey Caleb!

If you're working with prediction models for FRC, I'd suggest you checkout what my team, The Wired Wizards, did with ORB.

The link to the forum post is here: https://www.chiefdelphi.com/forums/s...d.php?t=147817

We worked with match prediction specific to Stronghold, doing analysis based on teams affinity with different defenses... there's more information in the post...

Good luck!

Did you ever see how accurate your predictions were? After training with data from 2011-2014, my current Elo model gets a Brier score of 0.1792 for predicting the winners of the 2016 Championship qual and playoff matches. I'm planning to try to predict matches with OPR, CCWM, and win contribution in the future and compare results.

[Caleb Sykes]

Originally Posted by Caleb Sykes

Did you ever see how accurate your predictions were? After training with data from 2011-2014, my current Elo model gets a Brier score of 0.1792 for predicting the winners of the 2016 Championship qual and playoff matches. I'm planning to try to predict matches with OPR, CCWM, and win contribution in the future and compare results.

Alternatively, it predicted 72.3% of the winning alliances correctly.

[weaversam8]

The project was exploratory, we were learning, but off the top of my head I think it was between 60 and 70 percent accurate.

[Chris is me]

Originally Posted by Caleb Sykes

Did you ever see how accurate your predictions were? After training with data from 2011-2014, my current Elo model gets a Brier score of 0.1792 for predicting the winners of the 2016 Championship qual and playoff matches. I'm planning to try to predict matches with OPR, CCWM, and win contribution in the future and compare results.

If you're going to compare to all of the various methods, would be cool to also compare to average qual score; anything less predictive than that is worthless basically.

[Caleb Sykes]

Originally Posted by Chris is me

If you're going to compare to all of the various methods, would be cool to also compare to average qual score; anything less predictive than that is worthless basically.

Do you mean each team's average qual score for the matches they have already played? Average qual score for the event would just predict a tie every match.

[Chris is me]

Originally Posted by Caleb Sykes

Do you mean each team's average qual score for the matches they have already played? Average qual score for the event would just predict a tie every match.

Yes, the average qual score of the team's previous matches; just like OPR uses the data from the team's previous qual matches, etc. OPR was always meant to be an improvement over average score. You can also do average winning margin as an alternate metric. OPR is supposed to be better than average score and CCWM is supposed to be better than average winning margin.

[Andrew Schreiber]

Originally Posted by Chris is me

If you're going to compare to all of the various methods, would be cool to also compare to average qual score; anything less predictive than that is worthless basically.

Wanna be lazier, try just predicting w/l based off existing win ratio for alliance at the event. I have a hunch it's as good as anything else.

[MattV781]

Waterloo 2016 Finals takes the cake for me. 1241 and 610 played incredibly well.