Loading...
  OR  Zero-K Name:    Password:   

WHR and the website are conspiring for higher accuracy!

8 posts, 965 views
Post comment
Filter:    Player:  
sort

2 years ago
I don't know how to start this so I'm posting my data.
I've cherrypicked these battles:
https://zero-k.info/Battles/Detail/1030889 59.3% chance of victory
https://zero-k.info/Battles/Detail/1031138 61.6% chance of victory
https://zero-k.info/Battles/Detail/1032676 37.5% chance of victory
https://zero-k.info/Battles/Detail/1023461 64.3% chance of victory
https://zero-k.info/Battles/Detail/1003640 59.5% chance of victory
https://zero-k.info/Battles/Detail/998843 33.9% chance of victory
https://zero-k.info/Battles/Detail/1001830 63.8% chance of victory
https://zero-k.info/Battles/Detail/1001819 58.9% chance of victory
https://zero-k.info/Battles/Detail/998854 62.3% chance of victory
https://zero-k.info/Battles/Detail/998847 32.5% chance of victory

I'm certain the balancer would have, at the time, claimed these were 50.3% and 49.7% prediction chances. But the website does not record the odds of victory at the time the game was played, instead it arrives at other numbers somehow. I'm mildly confused by this.

Interestingly, it seems to post-predict the games with accuracy, from this random dataset of cherry picked games where the victory chance is >58, the winners and losers seems to show up about the proper ammount(33% win chance wins 3/10 times)
+2 / -0

2 years ago
I do believe that the website percentage is calculated on the "current" WHR rather on the "as of then" WHR.

and since the dataset is what give the WHR the predicted value match the observed value is given.

I am not DDevs though, so I don't know if it's 100 correct explaination
+0 / -0
2 years ago
quote:
I do believe that the website percentage is calculated on the "current" WHR rather on the "as of then" WHR.


I also believe this is how it works.
+0 / -0

2 years ago
When calculating new ratings, WHR always calculates the the history of your ratings from your first to your last game. The predictions shown on the website are always based on the latest ratings, which includes the information from the games themselves. So the result from a new game is taken into account to update both your future and your past ratings. Then your updated past rating is used to predict that game.
+3 / -0
WHR has two-dimensional time: the internal x-time that means "prediction made for that moment in time" and y-time that means "prediction is made in that moment in time"!
+1 / -0


2 years ago
Perhaps the website should show the win chance at the time of the game? The required information is plaintext in the replay.
+2 / -0

2 years ago
Some of the games I saw in my history did show on the website about 50% win chances. A more comprehensive dataset would be better for interest. These matches I found by searching [A] Teams All Welcome with player counts above 6, then going into the past and picking the largest anomalies.
+0 / -0


2 years ago
Website could show both win chance at time of game AND win chance with current elo (as it does now). That way, if the win chance actually increases for the losing team over time, you know they really binned it!
+7 / -0