Methodology:
I am using data from 9 sources weighted equally:

• Dialidol busy signal %
• Dialidol raw votes
• Google mentions in the last week
• Posts in their subforum on idolforums
• Posts in their subforum on myidol
• Number of group members on myidol
• WhatNotToSing overall rating
• WhatNotToSing rating from their last performance
• Score from idoldemocracy

Yes, yes, it's ridiculous to weigh them equally -- I said stat-based, not scientific.

For each stat, I found the mean and standard deviation for all contestants with data combined (so including Lacey when she was around). I then determined the number of standard deviations each contestant was above and below the mean for each metric, and then averaged those to come up with an overall power score.

I now have 2 weeks of data, so I can make comparisons. If I'm not too lazy (so no guarantees), I'll try to update this each week around this time:

1. Crystal: 1.21 (LW: 1, 1.36)
2. Siobhan: 1.08 (LW: 2, 0.53)
3. Lee: .63 (LW: 4, .42)
4. Casey: .22 (LW: 3, .48)
5. Didi: .19 (LW: 5, .2)
6. Aaron: -.29 (LW: 9, -.39)
7. Katie: -.31 (LW: 10, -.5)
8. Michael: -.4 (LW: 7, -.4)
9. Tim: -.44 (LW: 6, .11)
10. Andrew: -.55 (LW: 8, -.35)
11. Paige: -.79 (LW: 12, -.98)

(Last week, Lacey would have been 11th)

Biggest change this week was Siobhan closing on Crystal (way to use data to support the obvious, pete!). Crystal had some regression to the mean going on, in part because she had a huge spike of google mentions after her diabetes scare that has been steadily dropping down to normal levels. Lee and Casey switched spots and then some, largely because of Lee's dialidol score and a huge increase in his forum traffic. My completely subjective guess would be that the teens that liked Alex Lambert are migrating to Lee instead of Casey. Didi, for all the buzz, had virtually no change, and Tim bombed across the board.

I don't know if I'll try to improve this totally sketchy model, but I will try to keep updating if y'all are interested.

This would be a very easy neural network to build over the rest of the season, fyi. cool idea.

wow, that's pretty cool...you must have a lot of time on your hands, lol

Not really. Yes, I'm procrastinating other things, but the data collection takes about 15 minutes and the spreadsheet took maybe 20 to set up.

I like this a lot, thanks for setting it up. I hope you keep updating.

Obviously two weeks of data isn't much to go on, but it certainly passes the eyeball test.

awesome, thxs for posting!
would be interesting how these ratios change.
as contestants get knocked out, i think the remaining singer pick up some of their votes.

also, maybe itues has some stats for downloaded songs or something? idk, but i think that could be a useful stat.

thanks for this pete

nice work

Yeah this is cool, and likely very accurate. I could see this being the order they get eliminated for sure.

Unfortunately, I don't seem to be on America's page this year. I think Mike, Tim, and Paige are all better contestants than Lee, Aaron, and Katie. I can see why people aren't voting for Tim and Paige -- but I do like them quite a bit more than your average viewer. Mike just perplexes me; why isn't he getting more popularity? He's likable, has good stage presence, sings well, and brings a unique style to the stage. Nobody else in the top 12 is remotely like him. It's weird that he's not doing as well.

Likewise, I just don't see the appeal in Lee and Aaron. Lee seems so far out of his depth as a contestant. He isn't ready to be a pop star and will just get swallowed up. I don't like his taste in music either. Aaron is a dope. Reminds me of one of those door-to-door Christian kids who push bibles in Utah. Voice isn't terrible but there's nothing appealing about him. As for Katie, she has a very good voice but is terribly unlikable on stage for some reason. She picks the wrong songs, performs weirdly, and doesn't have a real personality -- it all seems kinda fake.

Tim, on the other hand, is very down to earth, funny, and not at all phony. That's also probably why I like Mike, Didi, Paige, and Siobhan a lot too. Didi's personality does creep into that annoying realm every now and then but I give her slack because she's such a strong vocalist and has a lot of heart.

mike: he is "old" imo. a lot of black contestants are too conservative for AI imo. they sing stuff like R&B, and its hard for me to not forget it. i like mike, i just don't like the songs he sings.

Mike has kind of gone into lounge-singer mode the past couple weeks. I like him more when he was strumming his guitar and doing more fun songs.

I like the guy and he isn't going anywhere for a while, but he would have to really step it up to have any shot at actually winning.

As for Tim and Paige, well they are terrible and shouldn't have made the top 12. You are right about Aaron, he is awful as well and I don't get the popularity at all.

Random observation:

I don't have the data whipped into presentable form, but I've looked at the dialidol results for the past 3 seasons and found that the "dialidol score" is only marginally more accurate than the di raw vote count. And if you look at trends in the votes (i.e., adjust the results for who the DI user base tends to support and how they do relative to that baseline), it's actually even the better metric. If I get around to it, I may try to incorporate that into my model, but I'm a little less excited about AI analysis after this week's dog of an episode.

Stating the obvious here, but your model is going to underweight certain demographics that don't frequent internet fan forums, for example (which I suspect is one reason we haven't yet seen Andrew in the bottom 3).

Update (3/27):

1. Crystal: 1.58 (LW: 1, 1.21)
2. Siobhan: .70 (LW: 2, 1.08)
3. Casey: .36 (LW: 4, .22)
4. Lee: .30 (LW: 3, .63)
5. Didi: .04 (LW: 5, .19)
6. Aaron: -.12 (LW: 6, -.29)
7. Tim: -.27 (LW: 9, -.44)
8. Michael: -.37 (LW: 8, -.4)
9. Katie: -.42 (LW: 7, -.31)
10. Andrew: -.63 (LW: 10, -.55)

Crystal gains this week off of terrific Dialidol numbers. Siobhan drops, even out of the pimp spot. Lee also drops, unable to replicate his awesome DI numbers from last week. Tim makes a marginal gain that vaults him 2 spots even after a awful performance, though looking at the breakdown it appears to be largely from increased forum traffic (which could just be from a bunch of people debating how terrible he is).

Overall picture looks pretty similar, except with the gap between Crystal and Siobhan widening -- in fact, the gap this week between Crystal and Siobhan is larger than the one between Siobhan and Didi. Casey and Lee are still jockeying to be the top male, and no one else is really making noise, yet.

If you have any suggestions about weighting various factors, I coud try different things out. I have all the data in one place and pretty well organized. I've thought of trying a few things like including diminishing Last, Last 2, Last 3, Best, Best 2, Best 3 performances and that sort of thing. That's all easy enough to calculate, but picking particular weights seems so subjective. If it just ends up as "my theory of American Idol" then its value as data is kind of lost.

In my opinion, any ranking where Michael is below Tim, Aaron, and Didi cannot possibly we weighted properly.

I don't think I agree with this line of reasoning. The purpose of creating a model isn't to try to confirm pre-existing intuitions -- quite the opposite. Obviously if it produces some completely absurd result (like if it ranked Crystal last), it can be good evidence that there is a fundamental problem with the model that might make you start over from scratch. But even then, you have to be careful to make adjustments or design choices based on reasonable assumptions, not to produce your desired result. And in this case I hardly think the relative positions of Michael -- who has not posted great dialidol results, isn't getting great performance reviews, and doesn't have a lot of buzz on forums, etc -- to those other people (who are all fairly close to each other) is significant evidence of anything. I'm not saying my model is that good -- or really good at all -- but that's exactly the type of thing I would hope a good model would identify.

Again, if you have a substantive argument about what the weighting should be that's not based on how it would affect specific contestant's rankings, I'd be happy to hear them and maybe try them out.

Also, tangentially related, there are certain data points that I definitely do NOT want to include that are part of standard idol prognostication, like being the last black person standing, or being the only representative of a certain genre, or being a more favorable age, etc. That kind of stuff gets into subjective theories about how people vote and which traits are important, where I was more just interested in what we can see objectively from the unmitigated data -- which is also why I am wary to get too complicated with the weighting at all. My preference would be for broad and reasonably balanced, which may not lead to the most predictive model, but would give a reliably objective picture of where things stand and how the dynamic changes from week to week.

Obviously that comment was moronic. Didnt really need a long reply. Like this thread, though. You really should build (or provide the data to someone to build) a simple neural network though to figure out the optimal weighting of the different factors.

I know very little about neural network analysis. From my incredibly limited understanding, it's like regression analysis except it tries to find more complicated correlation functions?

The general idea is that the NN finds the best fit for the data that you have and then as you add more data it 'learns' what would be a better fit for the new data.

Here's a very quick but decent summary of the connection between NNs and statistics:

http://www.faqs.org/faqs/ai-faq/neur...ection-15.html

You probably could even find code that implements a simple NN on the web somewhere. Obviously you also could just run some simple regressions to find decent weights and re-run the regression every week...