The Top 3 tracker is a long view of the contest, trying to figure out who will end up in the Top 3 based on the current standings. It works in a similar way to the finals model, but rather than week-to-week, it looks at historical information to decide who most looks like a Top 3 contestant.
Note that these are all calculated before last night’s performances, so these are based on last week’s scores.
All things being equal, everyone’s chances should improve every week, as any time someone is eliminated you have a better chance just by random pick.
Tyanna, Clark, Nick, and Joey look like the most likely by a lot. All of them are better than 50:50. Hit the jump for more info.
Now we look at the Top 3 tracker over time:
Tyanna, Clark, Nick, and Joey all got over the initial hump. Nearly everyone else fell considerably, which is what usually happens. At first it’s hard to tell who will be a front-runner, but after a few rounds the herd separates. The only contestant who gained was Daniel, and even then only marginally.
To visualize the historical data, I have plotted as points the model safe probability versus whether or not the person made the Top 3. If the person is near the top, he or she made the Top 3, and if that person is near the bottom, he or she did not make it. Some vertical jitter has been added so that points can be easily seen. Mouse over a dot to see which person in Idol history it corresponds to.
(Note that Nick and Clark have overlapping points.)
It’s early yet, but none of these people would be shocking omissions from the Top 3. Nobody has ever been in the Top 3 with a lower score than Rayvon, and only 1 person has ever made it with scores worse than Maddie and Qaasim. Three people have made it with worse prospects than Daniel, including eventual winner Lee DeWyze.
The black curve is the theoretical probability of making it versus how likely someone was to be safe in the Top 11. At this point, it’s still somewhat uncertain who will make it, and the Top 3 tracker hasn’t yet been able to start averaging the data to get a better picture (averaging the data yields no improvement at this point). As the season goes on, the system becomes more sure of itself, as averaging the data gives a much better fit.
It is possible (I have not yet checked), that assigned probabilities could be better omitting song order from the fit. After all, the goodness of the contestant in theory has nothing to do with when they sang, although in reality the producers probably put their favorites toward the end. I will check this and post later.