Quote:
Originally Posted by ganstaman
Would it be reasonable to exclude the last round and see if that changes your results, or is that too much extra work?
That depends on how my data gathering goes, and more to the point how the data ends up being formatted. If the formatting allows for it to be checked easily then I would do so.
Quote:
Originally Posted by RoundTower
I think it's clear that live ratings carry more information and should be better predictors than pre-tournament ratings. And we also know that the fide k-factor is conservative and a higher k would give live ratings more predictive power (this is not necessarily bad because ratings are not used only for predicting future results).
If there's a hot hand effect, though, you should overweight the recent results, move the rating used for the sim by more than it moved for the live rating - even more than it would move with a k-factor optimised for long term predictive power.
I think that's clear too, intuitively, but would feel better if I could prove it. Updating everyone's rating to reflect the most recent round before starting the next round is tedious enough that if it turns out intuition is wrong here, and if data says it doesn't help, then it would be nice to stop doing it
Plus it feels like the kind of study that once I have the data, could also be useful for other things too. After the all-decisive first round I looked a little at those odds, and then it occurred to me to wonder about the odds of a round with NO decisive results (all five games drawn) and from there I started thinking about whether there might be any kind of round by round grouping effects in draw rates.
In theory, if every game is truly independent, and if draws happen at a consistent rate of 56%, then the odds of all five games in a round being decisive are 1/61 and the odds of all five games in a round being drawn are 1/18. This means that in any given 10-player round robin (nine rounds, five games per round) there should be a 60% chance that EVERY round has at least one decisive game. Seeing all five games in a round be drawn should be rare.
Norway saw all draws in round 7 (event draw rate was 56%). Looking again at how it "feels" to be (rather than actual data) this doesn't seem that uncommon to me. Not as uncommon as it "should" be statistically, unless there's a phenomenon that "groups" draws into certain rounds (maybe later rounds?), which would have important implications for my simulation methods.
Another not statistically significant data points: Tbilisi Grand Prix had 11 rounds (6 games per round), so a round with all draws should have been even rarer if the draw rate was the same. First nine rounds: at least one decisive game every round, total draw rate 58%. Final two rounds: 12 draws. Counterpoint: there was at least one decisive game in all 12 rounds at Khanty-Mansiysk. It's frustratingly hard to know what "patterns" we think we see are real, and which ones are just our overactive imaginations. This is why I love data
Long story short, if there is an effect that increases the draw rate late in tournaments, building that into my model (along with the corresponding DECREASED draw rate in early rounds, since that's how averages work), it would be useful for me to identify and quantify. The data set for that is probably the same as the data set for my live ratings study. Plus who knows what other unrelated and unexpected discovery might come from the research. Or maybe I'm totally overthinking it all, and nothing useful will come from it at all, I am prone to that