Predicting American Idol – Part 2 April 23, 2008
For the second part of my study concerning whether tweets from Twitter could be used to predict the losing contestants of American Idol, I decided to wait until noon to make sure that there were enough new tweets to equal the amount used in the analysis last week. Fortunately, there were more than enough fresh, unique tweets discussing last night's episode.
As I did last time, I used Tweet Scan to search for tweets about each of the remaining contestants on American Idol, stopping at 90 each. Then, I ranked each of the tweets as positive, negative, or neutral, which is similar to the ImpactWatch system of rating articles.
Syesha Mercado had the most positive new tweets (78), despite the fact that she had the fewest just a day ago. David Cook continued to have mostly positive tweets, while Carly Smithson nearly doubled her amount. Contestants Jason Castro and Brooke White dipped in positive tweets, while their negative tweets rose dramatically. David Archuleta, who was the front-runner at the beginning of the season, continued to slide in positive tweets and gained more negative ones. An Excel spreadsheet of all the specific data is located at the end of the post, while a graphical analysis is below.
While it could be useful simply to take the most recent tweets into account, I decided that to make an accurate prediction, I should combine the previous night's tweets with the more current ones. It seems reasonable that voter opinions would last for about a week with the most current performance taking precedence in minds.
When looking at the combined data (graph below), Cook seems to be the frontrunner, with Mercado, Castro, Archuleta, and Smithson all behind him. White is the only contestant to have more negative tweets than positive tweets overall. If what I hypothesized is true, it seems from the data that the dreaded Bottom Three will contain White, Castro, and Archuleta, with White sealing her eliminated fate.
As I have said previously, I do not believe that online activity can predict offline behavior, but since many of the tweets specifically reference contestants for whom the writer had voted or was planning to vote, I believe that tweets may represent a fairly accurate portrayal of the voting population. We will see if I am correct tonight.
Below is an Excel spreadsheet of all the data collected during the course of the study. Also below is a report, created using ImpactWatch, showing all of the graphs in a high quality format.
