SubscribeTom Tango has started up this year’s Community Forecasts, a project that recruits hundreds of fans to predict playing time for players on their favorite teams.
I took a snapshot of the PA and IP from the voting as of this evening and added a new option to the Price Guide to use it for playing time. It’s the new default option for the Composite Projections — “2010 Composite (CF)”.
At first glance, this appears to be an improvement over my previous manually tweaked playing time values. Several of the players who seemed too high (Carlos Guillen, David Freese, Matt Diaz) look much more reasonable, now. Garrett Jones, Nyjer Morgan, and Franklin Gutierrez are among the biggest gainers.
The fans are also doing a better job of sorting out the Rockies OF and 2B situations that I was: Seth Smith, Carlos Gonzalez, and Clint Barmes each gain a few dollars. Brad Hawpe, Eric Young, and Dexter Fowler drop a little bit.
Let me know what you think. I plan to keep updating periodically as the voting continues.
Related posts:
Not a bad idea.
Would you be able to announce when you update the values so we know when changes have been made?
looks like clement took a huge fall as well
ps I think this is the best method for determining PT yet – very helpful even though I already had my draft!
@Dave: That’s a good idea. I’ll either post here or set some kind of “last updated” timestamp on the Price Guide.
@John K: Yeah, I agree these values look much better than past iterations. Even if it ruins so many of my previously declared sleepers, like Clement.
Awesome stuff as always. I am using these as my last, final projections.
Actually, that’s not quite true. I have found that my league drafts somewhere in between your “Optimal Split” and your “70/30″ split. So I run my league settings each way, and take the average of the 2. THAT is what I use.
anyways, great stuff
One thing I’m not too comfortable with is that Tango claims he’s using the wisdom of crowds to get at good playing time forecasts. As I understand it though, submissions are voluntary and so there is likely a selection bias. Much more of an issue with something like Fangraphs’ fan stat projections than with PT guesses, but still not ideal.
if Tango is comfortable with it, so am I
John, the proof is in the pudding: fan forecasts are almost as good as anything out there.
Any chance you can make the composite projections available? I think the Price Guide is great, and a big chunk of its value is in the quality of the projections behind it. I’m playing in a points league and would like to come up with my own values, but would love to have your composite projections (with the community PT). Thanks for the good work with the Price Guide.
Harold, you can just run the Price Guide, then go to the League Info tab. There is an Export to CSV link there. Use that and you get the composite projections with the community playing time in your CSV (or whatever projections you chose to build the price guide with).
Tom: Thanks for the response – I haven’t done any work around forecast evaluation so that is good to know. Like I said, I believe the selection bias is a bigger issue w/performance based stat projections conditional on PA or AB.
I would not be surprised to learn that fan projections are better for PT stats, but your statement seems more broad than that. As I mentioned above I have a favorable disposition to polling individuals for projections in general, but I think the statistics would be greatly improved if there was some mechanism to ensure a diverse contribution.
For instance, I looked at fan projections vs. Marcel and Chone and set a simple criterion to test my theory that fan projections will be more effusive: A binomial test that a player’s OBP is greater than a projection algorithm yields the following:
vs Marcel
number of successes = 240, number of trials = 366, p-value = 2.619e-09
vs Chone
number of successes = 261, number of trials = 358, p-value < 2.2e-16
vs simple M&C average
number of successes = 259, number of trials = 358, p-value < 2.2e-16
mean tests across the distributions also reject the null of equality.
I would say it looks like the fans are being too optimistic. Do the projection algorithms systematically under-predict the rate at which batters reach base safely (I would have expected not)? I don't know the answer to that type of question, but just something I was thinking about – sure it's not the first time it's been brought up.
*I also did one-sided tests of other stats like HR/AB and found the same thing (especially with Marcel, which apparently predicts much less power)