Nate Silver got a gig with arguably the most prestigious paper in the US, the New York Times. I think that's why he got hyped up so much more than others following similar methodology with similar results.
That's probably it. He's been doing it for a shorter time, and has benefited from those who started four years earlier than him, and still seems no more accurate than other who are using the same methods. His graphics are certainly pretty.
@Luke: Nate Silver's data is more interesting, and it's presented in a much more pleasant way. Take a look at the right side of his site - Tipping point states, ROI index, scenario analysis - this is all really interesting information. His blog posts are insightful and informative. Also, his methodology actually does differ from his competition in some important ways. From his FAQ:
How is this site different from other compilations of polls like Real Clear Politics?
There are several principal ways that the FiveThityEight methodology differs from other poll compilations:
Firstly, we assign each poll a weighting based on that pollster's historical track record, the poll's sample size, and the recentness of the poll. More reliable polls are weighted more heavily in our averages.
Secondly, we include a regression estimate based on the demographics in each state among our 'polls', which helps to account for outlier polls and to keep the polling in its proper context.
Thirdly, we use an inferential process to compute a rolling trendline that allows us to adjust results in states that have not been polled recently and make them ‘current’.
Fourthly, we simulate the election 10,000 times for each site update in order to provide a probabilistic assessment of electoral outcomes based on a historical analysis of polling data since 1952. The simulation further accounts for the fact that similar states are likely to move together, e.g. future polling movement in states like Michigan and Ohio, or North and South Carolina, is likely to be in the same direction.
@trogdor: I agree that his graphs are pretty. To be honest, I don't see any small change in methodology to be that important when making those changes makes you no less or more correct than others using the same basic method.
I bring this up because multiple times I've seen posts about Nate Silver, saying he is awesome, and then a link to a dartboard where the closest dart to the center (actually dead center) is another analyst who did the same kind of modeling. If pure accuracy is what you want, Silver isn't the best (and calling two elections in a row is like flipping a coin twice, a quarter of those who do it will get heads twice). If being first is what you want, he wasn't first.
The reason I've enjoyed looking at other similar sites is that they outline their exact process, and invite you to check their methods and numbers and results. I can't find the same invitation to examination on 538.
I think, maybe, that hiding the process, and by adding minor complexity that doesn't noticeably affect the predicted result, Nate Silver seems more mysterious and magical, when really he's not adding all that much.
Silver's popularity comes from three sides, Luke. First Silvers was already popular as a statistician before entering politics. He is the guy who developed PECOTA, a statistical system to asses and forecast the performance of baseball players.
The second prong of his popularity is accuracy. And that is not just two presidential elections, but he also has made rather accurate predictions for congressional and gubernatorial races in the U.S. for the last couple of election cycles, plus his history in Baseball predictions.
The third and most recent prong is the blatant antagonism he has drawn from the political right, who tried to discredit him because his prediction is rather trusted. If they succeeded with it they hoped to discredit all projection models which disfavor them by proxy and thus mitigate the "self-fulfilling prophecy" effect that such statistical models can potentially have. Silver stood up to that bullying, and as it shows, he had he was mostly correct in his predictions.
So in sum total, Silver's method may or may not be the best and most accurate prediction model out there, but that isn't the only factor in his popularity or necessarily the reason why people are going "gaga" over him.
While it's true that it's nice to have transparent methodology, I'm inclined to ask what it is you're saying others did better than Nate Silver.
Who and what exactly are you referring to when you speak of "dead center" or "first"?
First, Dead Center.
I saw a few tweets link to this page, all commenting about how amazing Nate Silver was: Slate's Pundit Scorecard.
The important part of this story isn't that I'm saying Nate Silver isn't accurate, or that his methods aren't clever, but that I was surprised to find, when looking at the Pundarts section, that Nate Silver wasn't the most accurate. Josh Putnam, of FrontloadingHQ, had actually hit the target dead on. And yet, at the time, all these tweeters ignored someone with a more accurate prediction, and were congratulating someone with a less accurate prediction.
The page has now been updated: "Correction, Nov. 7, 2012: ... Nate Silver's prediction had previously been listed as 313 electoral votes. This number was Silver's average of simulated outcomes, not the most likely outcome." I admit that this incorrect reporting of Nate's prediction means he was equally accurate as Josh Putnam (exactly correct), but at the time I saw the tweets, the Nate Silver fans seem to have ignored the fact that someone else was either equally accurate or more accurate, and there was no awareness of that.
Second, First.
Election Projection was set up for the 2004 election. Electoral Vote was set up for the 2004 election. I'm sure there were others, and there have been many more since (like FrontloadingHQ mentioned above).
And when I say "learning form those who went before", the Electoral Vote Votemaster is open about his methods, and in 2004 gave predictions based on three models. From wikipedia:
"The site's final tracking using algorithm 1 posted on Election Day, November 2 gave 262 electoral votes to John Kerry and 261 to George W. Bush, with 15 tossups.[2] The second algorithm (averaging 3 days worth of nonpartisan polls) gave Kerry 245 and Bush 278 with 15 tossups. The third algorithm (predicting the undecideds) predicted 281 for Kerry and 257 for Bush.
The actual vote gave Kerry 252 to Bush's 286. Using nonpartisan polls and averaging a few days worth of polls did best."
If you're going to start four years later than others doing a very similar kind of prediction, this kind of information is pretty handy.
Again, I'm not saying Nate Silver isn't accurate so far, or that his website doesn't look very pretty, or that his graphs are very informative, but that I'm quite bemused about the adulation I'm seeing around the internet. He really isn't a witch. Is he?
I have more to say about election predictions, and maybe why Nate Silver is a genius. There's usually two ways to predict the election: either by who wins, or by how many electoral votes they get.
Those who just say who wins are making a binary-style statement. A or B. Calling the race one way or the other is something that anyone can do, and anyone with an opinion can simply say it out loud. That's a pundit prediction. And if you are wrong you can bitch about the popular vote, etc.
The other way, to predict the electoral vote, is a more complex beast. It isn't a single binary statement, but the sum of 51 binary statements, each with their own weight.
This means that the Pundart score board is bullshit. The three darts in the center (as of this writing) actually got a score of 51/51.
The first dart out from the center (Donna Brazile) is only "-19" off, which on the face of it makes it look like she is "more accurate" than the next dart (Sam Wang) at "-29". However, this actually means that Donna got 49/51 of the state votes right, and Sam Wang got 50/51.
Everyone who predicted Florida would go to Romney put Obama's electoral vote at 303, not 332, and likewise got a 50/51 result. And this result (Barack Obama: 4,143,362, Mitt Romney: 4,096,346) is crazily close, though admittedly not as close as it has been in the past.
Anyway, what I think Nate Silver is doing differently (though the last forecast I can find on his website is still the 49/51 "Updated 10:10 AM ET on Nov. 6" one) is setting the overall "Chance of Winning" so prominently (90.1%), and tied directly to the electoral vote prediction. Normally the chance of winning is tied to the first kind of prediction, the A or B kind, or to the popular vote. Nate puts it above the popular vote.
However, Nate Silver has presented his result so that the "confidence" rating is waaaaaay broader than the final electoral vote could ever have made the race. EVERY aggregated state-level polling analyst (all the ones I can find) had the same confidence that their model would predict the winner *in this election*, and EVERY ONE was right. This is so obvious that there is even an xkcd about it.
Nate's own information on his own site, in a very clever move, puts the "importance" (probability that a state provides the decisive electoral vote) of Florida's result in 9th place, at just 1.2%. This means that, in Nate's own analysis, calling the correct result in Florida was only 1.2% of his final prediction.
Which, in turn, means that all the other analysts using similar methods, and called Florida incorrectly (understandably when there was fewer than 50,000 in it), were not only 50/51 compared to 51/51 for Silver (though I still can't find his correct predicition on his site), but that the one state they didn't call was only 1.2% significant.
Nate Silver, by having his "uncertainty" so prominently, could easily have got away with calling Florida incorrectly. He states his overall uncertainty above the much smaller number of 1.2%, and this can also overwhelm any difference in the electoral vote.
On the bottom of this page there is a link "Now here's why Nate Silver got it right". And what does that page tell us? Not why, just "that" he got it right. And it's an EPIC Victory for Nate Silver. Really? Epic?
Then again, I'm a big fan of certain bands, and think they have EPIC new albums and EPIC stage shows. I guess, for nerds, Nate Silver is just the star statistician, and I'm the confused onlooker thinking "all he has are pretty graphs and a name".
Again, I'm not saying Nate Silver isn't accurate so far, or that his website doesn't look very pretty, or that his graphs are very informative, but that I'm quite bemused about the adulation I'm seeing around the internet. He really isn't a witch. Is he?
Yes, Nate Silver is getting an undeserved amount of credit compared to others who have done much the same thing. Mostly I guess it's just the usual situation, with a good boost due to the popularity of the New York Times followed by the type of popularity explosion that is all too typical of the Internet; this is coupled with the effects of people being bad at understanding statistics.
In short, people are surprised to see that statistics is a thing that actually applies to the real world, and Nate Silver happened to have several factors that put him at the crest of the popularity wave.
That being said, his site is, overall, the best of the bunch; the reasons for this include those already mentioned - good site design and informative graphs. The blog posts are also quite good, and they do in fact give good details on his methodology.
Personally, I value Silver's use of direct and easily accessible probabilistic estimates. "My model assigns an 80% chance to Obama winning Virginia" is much more informative than simply saying they expect Obama to win Virginia, and in particular the clear and accessible probability of overall victory (90.9% chance of Obama victory) is highly convenient.
Sam Wang's PEC is OK, and strictly in terms of accuracy the 99.2% chance of Obama victory gets them more points than they lose for getting Florida wrong, but despite their being right 99.2% still seems overconfident to me.
On the other hand, though, it looks like Nate Silver's Senate predictions definitely weren't as good as others, especially his prediction of 92.5% for Republicans to win the North Dakota Senate (which they did not) - as compared to Sam Wang's 75% for Dems to win. Pollster doesn't give probabilities, but it would appear that they, and Sam Wang, correctly predicted all of the Senate seats.
This means that the Pundart score board is bullshit. The three darts in the center (as of this writing) actually got a score of 51/51.
The first dart out from the center (Donna Brazile) is only "-19" off, which on the face of it makes it look like she is "more accurate" than the next dart (Sam Wang) at "-29". However, this actually means that Donna got 49/51 of the state votes right, and Sam Wang got 50/51.
Both of these are terrible ways of scoring, and poor indicators of predictive power.
Nate's own information on his own site, in a very clever move, puts the "importance" (probability that a state provides the decisive electoral vote) of Florida's result in 9th place, at just 1.2%. This means that, in Nate's own analysis, calling the correct result in Florida was only 1.2% of his final prediction.
No, this is totally wrong. The probability of providing the "decisive electoral vote" is very, very different to the "importance" of the state. This is something that I think FiveThirtyEight is a little bit misleading on, but they do a decent job of explaining what they mean by the term here: http://fivethirtyeight.blogs.nytimes.com/2012/04/27/arizona-is-probably-not-a-swing-state/
Something that I think needs to be made more clear with regards to which state is the "tipping point" is that it's a matter of correlation, not causation. Despite everything, Ohio is still worth only 16 electoral college votes; Romney winning Ohio would be a big deal only because it would be correlated with Romney victories in other states.
That's a good roundup of the article, though I'm not sure of all the reporting.
Here's an excerpt "[Silver] considered Florida a virtual toss-up, with Mr. Obama and Mr. Romney tied at 49.8 per cent apiece, but he gave the Democrats a 50.3 per cent chance of taking the state."
That's just an incredibly small number! 0.6% is probably within the margins of error of his own prediction, and this is a prediction of the chance that Obama would win, not even his prediction of the vote count in the state. Which, if it was that close, would have been cause for a recount anyway, which means that Silver's prediction was inside the margin of error for the election itself. That's not good science!
Here's another thing. Nate Silver's page is tricky to understand, and it's difficult to tie down his actual prediction. The top number is the 313 electoral votes for Obama, and this is the one that was chosen by the Slate Pundarts graphic. This prediction was incorrect.
However, scrolling down you find the State by State probabilities map which, if you just go by the shading of the states, is exactly the same as the final state results. However, the shading doesn't show how "strong" the state is for either candidate, like other similar prediction sites (who use poll results to show what the vote count might be), but the likelihood that a candidate will win. This is confusing, and belies the fact that the 50%-60% likelihood that Obama would win Florida way, way broader than, say Electoral-Vote's outlined red shading for Florida, which shows that Romney might get between 1% and 5% more votes than Obama.
What I'm trying to say is that, by not declaring ties, and using 10% ranges for probabilities of winning, rather than much smaller ranges of total state votes, Nate Silver's maps are actually less accurate than other sites. His prediction might have been more correct, but his margin for error was far larger, and only delineated by the 50% cutoff line.
I'm not the only one to be confused by Nate Silver's presentation of his prediction, as the Slate Graphic also was confused about his actual prediction.
The same difficulty of reporting is shown by the linked article in regards to the Electoral-Vote site. They say "The poorest performer out of the more well-known forecasting websites, Electoral-Vote.com, only missed the call in Florida and considered North Carolina a tie (it was won by Mitt Romney). "
The Electoral Vote website didn't think any state would have a tied election. To predict that a state would be exactly split between two candidates would be absolutely bonkers. What the website's map actually showed was that the result was within the margins of error with the polls it was using for its calculation. I think this kind of reporting is actually far more interesting than the "0.3% over 50-50 means I'm calling it for Florida" that Nate Silver's prediction makes. I liked that there were those 15 electoral college votes up for grabs in North Carolina.
The website even says "It is worth emphasizing that the margin of error in most state polls is at least 3% for each candidate... Even some of the ones in the solid light color may technically be statistical ties, but a lead of 5% or more most likely means the candidate is actually ahead."
Finally, Nate Silver has a graph of the probabilities of the actual electoral vote count: "The probability that President Obama receives a given number of Electoral College votes."
332 is at about 20%. 303 is at about 16%. That is a 4% difference between Nate's prediction and the many others who predicted 303. What this shows is that Nate was only 4% more confident in (one of his) actual predictions (the one that you have to infer from the map) over the next most likely (the prediction that many other sites made, and was well within their stated margins of error). His stated number of 313 is hard to find on that graph at all, and so it's unclear how that number is listed at the top of his November 6th prediction.
Lessons learned:
Put at least two different predictions on your website, and when the result comes in, let people pick the one they like best for reporting. If they are a fan, they'll pick the less-accurate but more correct map. If they aren't so sure, they'll pick the top number, and report that you were very close. Either way, you're no worse off than the other sites who were more upfront with their margins of error and less correct.
It really all comes down to Florida, which Nate shows as:
Now-cast 48.2 48.2 Tie Projected vote share ±2.7 49.8 49.8 Tie Chance of winning 50% 50% .
All those numbers are the same. He made a call 0.3% "certainty" call for Obama, the popular vote was larger, and he looks like a genius to those who haven't looked closely at his numbers.
Here's an excerpt "[Silver] considered Florida a virtual toss-up, with Mr. Obama and Mr. Romney tied at 49.8 per cent apiece, but he gave the Democrats a 50.3 per cent chance of taking the state."
That's just an incredibly small number! 0.6% is probably within the margins of error of his own prediction, and this is a prediction of the chance that Obama would win, not even his prediction of the vote count in the state. Which, if it was that close, would have been cause for a recount anyway, which means that Silver's prediction was inside the margin of error for the election itself. That's not good science!
The only bad science would be for Nate to give his model significant credit for "calling Florida"; Nate did not do this.
Finally, Nate Silver has a graph of the probabilities of the actual electoral vote count: "The probability that President Obama receives a given number of Electoral College votes."
332 is at about 20%. 303 is at about 16%. That is a 4% difference between Nate's prediction and the many others who predicted 303. What this shows is that Nate was only 4% more confident in (one of his) actual predictions (the one that you have to infer from the map) over the next most likely (the prediction that many other sites made, and was well within their stated margins of error). His stated number of 313 is hard to find on that graph at all, and so it's unclear how that number is listed at the top of his November 6th prediction.
Yes, and it's valuable information that Nate had ~20% probability for 332, and ~16% probability for 303. The fact that this type of information is missing from many other sites is to Nate's credit, and to the detriment of those other sites.
Pollster and PEC did give graphs like this, though, so let's have a look at the graphs:
Simon Jackman for HuffPost Pollster:
Sam Wang for Princeton Election Consortium: (I think the red line is a glitch in the graph here)
Nate Silver for FiveThirtyEight:
Pollster did slightly better here as they assigned ~23% to the actual outcome, and indeed from many other aspects Pollster has indeed been shown to be the most accurate site overall. However, it's ridiculous to criticize FiveThirtyEight for only being 20% confident of an outcome of 332 electoral votes as compared to other sites that don't even assign a confidence number at all.
However, scrolling down you find the State by State probabilities map which, if you just go by the shading of the states, is exactly the same as the final state results. However, the shading doesn't show how "strong" the state is for either candidate, like other similar prediction sites (who use poll results to show what the vote count might be), but the likelihood that a candidate will win. This is confusing, and belies the fact that the 50%-60% likelihood that Obama would win Florida way, way broader than, say Electoral-Vote's outlined red shading for Florida, which shows that Romney might get between 1% and 5% more votes than Obama.
What I'm trying to say is that, by not declaring ties, and using 10% ranges for probabilities of winning, rather than much smaller ranges of total state votes, Nate Silver's maps are actually less accurate than other sites. His prediction might have been more correct, but his margin for error was far larger, and only delineated by the 50% cutoff line.
The "10% ranges" you're talking about are not representative of any kind of margin of error; they're just simplifying aspects of the color scheme. Hovering over the state in question gives you the specific probabilities, like 50.3% for Florida. Your justification for Nate's map being less accurate than other sites is simply wrong.
In simple terms, everything that isn't the darkest shade of blue or red on Nate's map is a case that falls within the model's overall margin for error; the probability numbers indicate the proportion of projected outcomes that give Obama the state vs. giving Romney the state. Considering that the only relevant outcomes are whether Obama or Romney wins a given state, this is a concise representation of the important information at hand.
In effect, Electoral-Vote (which you seem to like) is wasting a lot of resolution on states that aren't really in play. Coloring Minnesota light blue is simply a waste of resolution that's better used on the more interesting states - having the same color scheme for Pennsylvania as you do for Colorado is ridiculous, given that Obama pretty much had Pennsylvania in the bag.
That said, I think the best approach would be to use a continuous color scheme rather than a discrete one. This would in all respects be a better way to present the information at hand, and in such a color scheme Florida would have been white and not red or blue.
The same difficulty of reporting is shown by the linked article in regards to the Electoral-Vote site. They say "The poorest performer out of the more well-known forecasting websites, Electoral-Vote.com, only missed the call in Florida and considered North Carolina a tie (it was won by Mitt Romney). "
The Electoral Vote website didn't think any state would have a tied election. To predict that a state would be exactly split between two candidates would be absolutely bonkers. What the website's map actually showed was that the result was within the margins of error with the polls it was using for its calculation. I think this kind of reporting is actually far more interesting than the "0.3% over 50-50 means I'm calling it for Florida" that Nate Silver's prediction makes. I liked that there were those 15 electoral college votes up for grabs in North Carolina.
The website even says "It is worth emphasizing that the margin of error in most state polls is at least 3% for each candidate... Even some of the ones in the solid light color may technically be statistical ties, but a lead of 5% or more most likely means the candidate is actually ahead."
Clearly no one thinks that Electoral-Vote expected NC to be a tie, but although Obama had decent chances of winning NC based on the available data (as was recognised by FiveThirtyEight and Pollster), it wasn't really a tossup by any reasonable metric.
It appears that Electoral-Vote's choice to shade North Carolina in white was based on a very limited number of polls. By comparison, FiveThirtyEight had North Carolina at 50.6%-48.9% with a 2.6% margin of error, which is indeed something that could go either way but is nonetheless solidly in Romney's favour.
Nate Silver's website does not in any way say "0.3% over 50-50 means I'm calling it for Florida". The probability of ~50% means that the state could very easily go in either direction. It seems to me that your only real complaint is about the color scheme.
Here's an excerpt "[Silver] considered Florida a virtual toss-up, with Mr. Obama and Mr. Romney tied at 49.8 per cent apiece, but he gave the Democrats a 50.3 per cent chance of taking the state."
That's just an incredibly small number! 0.6% is probably within the margins of error of his own prediction, and this is a prediction of the chance that Obama would win, not even his prediction of the vote count in the state. Which, if it was that close, would have been cause for a recount anyway, which means that Silver's prediction was inside the margin of error for the election itself. That's not good science!
The only bad science would be for Nate to give his model significant credit for "calling Florida"; Nate did not do this.
I think the point that you're not getting from all of my comments is not that Nate Silver is doing bad science. I really don't think so at all! I think it's really amazing that so many different sites and people and teams are doing this kind of analysis and prediction.
All I'm doing, with all of my critique (critique, not criticism), is trying to work out why so many people have gone gaga over Nate Silver compare to other similar sites and analysts.
My point about this map is that, due to his colouring scheme, where a light blue range can cover a predicted outcome 10% wide, and a result that falls within the margin of error (which 0.3% certainly qualifies for) is not distinguished. On another site, this same kind of result is given a white or outlined colour.
What happens is that, when someone looks at Nate's state by state prediction map, and it matches up with the actual results, they are hailing Nate as a genius, DESPITE Nate himself not calling the state of Florida, DESPITE him labeling it a "Tossup", DESPITE him giving polling numbers that show it an exact tie, and DESPITE him predicting an exact 50%-50% chance of it going either way.
And this where I think Nate was really lucky with his chosen algorithm. Even with all of his probabilities, and his top line prediction of 313, and all the rest, it was the way he decided to shade that one map, which compared with similar maps shows one thing (comparative votes) and with his shows something else (probabilities) that made so many people call him a genius.
I'm not trying to discredit Nate, I'm really not. I'm just trying to work out why myself, Slate.com, and many others looked at his site and saw a very good but not correct prediction and thought "Well, that's safely within the margin of error", and why all the people I've seen on twitter and blogs look at the SAME page and say "Wow, this guy made EPIC predictions and is clearly a witch!"
lackofcheese is obviously way more informed about this than I, so I don't really have much to add.
Sam Wang for Princeton Election Consortium: (I think the red line is a glitch in the graph here)
The red line is at 270 electoral votes, the number required for an election. To the left of that line means Romney wins; to the right means Obama wins. I don't really think the idea of Nate Silver having prettier/easier to understand graphs needs much more discussion, but it's definitely worth pointing out that Wang's poor chart design caused you to misunderstand what he was saying.
Nate's own information on his own site, in a very clever move, puts the "importance" (probability that a state provides the decisive electoral vote) of Florida's result in 9th place, at just 1.2%. This means that, in Nate's own analysis, calling the correct result in Florida was only 1.2% of his final prediction.
No, this is totally wrong. The probability of providing the "decisive electoral vote" is very, very different to the "importance" of the state. This is something that I think FiveThirtyEight is a little bit misleading on, but they do a decent job of explaining what they mean by the term here: http://fivethirtyeight.blogs.nytimes.com/2012/04/27/arizona-is-probably-not-a-swing-state/
Something that I think needs to be made more clear with regards to which state is the "tipping point" is that it's a matter of correlation, not causation. Despite everything, Ohio is still worth only 16 electoral college votes; Romney winning Ohio would be a big deal only because it would be correlated with Romney victories in other states.
Okay, that was a mistake. I'm not sure I'm going to read back to a blog post in April to clarify that though, and I think shows how Nate's long-term readers (and detractors) get one thing from his site, and people like me don't see what the fuss is about.
But my broader point stands. Many sites called the election on exactly the same terms. All of them showed that, even with Florida as a tossup, the election wouldn't be decided there. My point is that if, like in 2000, the decision came down to the Florida vote (a 1.9% chance?), Nate Silver was only putting it in the blue by 0.3%. The reason why, up until the election, nobody really cared about this tiny tipping of the scales, well within all margins (of polling and voting), was that Florida just didn't matter so much this time out.
Other states calling put Romney ahead there. It didn't matter for the final result. With the final votes under 50,000, it doesn't matter at all on a national level in terms of popular vote, and it affects nothing in the state itself except as a slight shading on a map. And Nate got that shading right! Along with others.
Yes, and it's valuable information that Nate had ~20% probability for 332, and ~16% probability for 303. The fact that this type of information is missing from many other sites is to Nate's credit, and to the detriment of those other sites.
Yeah, it's a really good metric. I think it shows a far clearer picture than other kinds of graph, especially the shaded map, which is very "static", if you know what I mean. The map belies a confidence in the result that the data and analysis doesn't actual bear. A 4% greater likelihood of one electoral result over another is not that great at all. It's actually impressive that all of the graphs are so similar. I'd love to see the three predictions on one graph with the same scales, to see how they differed.
I think that other sites aren't so concerned about forecasting as others though. For example, Electoral Vote (which I'm using an example because I've been following it since 2004, and know the approach the best), while calling itself a vote predictor, is less of a forecasting website and more of a poll tracking website. The map it presents is always of the data within the most recent polls, averaged over a certain number of days. It doesn't take into effect momentum or demographic changes (though those are topics within the blog).
It also doesn't assign weight to different polls and pollsters, and readily admits that bad polling (though declared non-partisan polls) will affect its final score. The votemaster does, of course, acknowledge that certain polls have a bias, and presents a Rasmussen-free results map. And, if you don't like the bare results of it presents, you're free to download the data it uses, every day, and do your own analysis.
Also the votemaster is a non-USA resident libertarian who lives in the Netherlands. Nate obviously has a personality amenable to either demonizing or deifying, depending on the outlook of the journalist involved.
It's down to what you want. I like a more transparent, "this is our data, our method, and our results" approach, and Nate has a "I have secret sauce, and run thousands of simulations, and have all these graphs of uncertainty" approach, which comes off as more magical, or at least more mystical. And when you want to create a modern myth, transparency and sticking with pure data is the opposite of what you need.
My point is that if, like in 2000, the decision came down to the Florida vote (a 1.9% chance?), Nate Silver was only putting it in the blue by 0.3%.
That's not really true, because the prediction conditional on the decision coming down to Florida is quite different to the overall prediction. If it did come down to Florida, that would mean that Romney must have made big gains in other states and hence would have much better odds of winning Florida than initially expected.
I'm not trying to discredit Nate, I'm really not. I'm just trying to work out why myself, Slate.com, and many others looked at his site and saw a very good but not correct prediction and thought "Well, that's safely within the margin of error", and why all the people I've seen on twitter and blogs look at the SAME page and say "Wow, this guy made EPIC predictions and is clearly a witch!"
What do you mean by "very good but not correct"?
All I'm doing, with all of my critique (critique, not criticism), is trying to work out why so many people have gone gaga over Nate Silver compare to other similar sites and analysts.
The biggest reason is also the simplest one - people don't understand statistics. To the vast majority of people, (sufficiently advanced) statistics is indistinguishable from magic, and hence anyone using statistics is a witch.
My point about this map is that, due to his colouring scheme, where a light blue range can cover a predicted outcome 10% wide, and a result that falls within the margin of error (which 0.3% certainly qualifies for) is not distinguished. On another site, this same kind of result is given a white or outlined colour.
What happens is that, when someone looks at Nate's state by state prediction map, and it matches up with the actual results, they are hailing Nate as a genius, DESPITE Nate himself not calling the state of Florida, DESPITE him labeling it a "Tossup", DESPITE him giving polling numbers that show it an exact tie, and DESPITE him predicting an exact 50%-50% chance of it going either way.
And this where I think Nate was really lucky with his chosen algorithm. Even with all of his probabilities, and his top line prediction of 313, and all the rest, it was the way he decided to shade that one map, which compared with similar maps shows one thing (comparative votes) and with his shows something else (probabilities) that made so many people call him a genius.
I agree with some of what you're saying here, but the issue isn't one of probabilities vs comparative votes and error margins - the probabilities are simply a more concise representation of the relevant information.
The real issue is that people are too obsessed with the idea of "calling" states, and Nate's color scheme doesn't help in this regard. It would be better if it was colored on a continuous spectrum according to the probability, rather than split up into discrete categories - this would have caused Florida to be white on Nate's map, with an undetectable blue tinge.
Taking a pause from the "Is Silver making just very good or the best predictions" discussion, here's an article from the Sports website Deadspin (yeah, it's Gawker, but whatevs) about Nate Silver, his history in baseball, and the similarity of the criticism that Silver's models received both in baseball and in politics.
My point is that if, like in 2000, the decision came down to the Florida vote (a 1.9% chance?), Nate Silver was only putting it in the blue by 0.3%.
That's not really true, because the prediction conditional on the decision coming down to Florida is quite different to the overall prediction. If it did come down to Florida, that would mean that Romney must have had to have made big gains in other states and hence would have much better odds of winning Florida than initially expected.
Right there on the very map I'm talking about (which has nothing to do with the tipping point rating of Florida) has this:
"State by State Probabilities: Florida Nov. 6th Forecast: 50.3%"
You have to mouse-over the state to see it. Scrolling down, to see more information about Florida, you find all the polling data to be equal:
Florida FiveThirtyEight Projections Dem Rep Margin Polling average 47.6 48.3 Romney +0.7 Adjusted polling average 48.3 48.1 Obama +0.2 State fundamentals 46.2 49.3 Romney +3.1 Now-cast 48.2 48.2 Tie Projected vote share ±2.7 49.8 49.8 Tie Chance of winning 50% 50%
For some reason, which I can't find by looking at the page, there is a discrepancy between the data in that forecast and the probabilities map.
The tipping point rating and percentage has nothing to do with the state by state probabilities, and I'm obviously not being clear in my writing about them. I think, however, this lack of clarity in Nate Silver's presentation is what leads people to see things in his predictions that he himself either isn't stating, or is stating with far less confidence than people are taking away from it.
For some reason, which I can't find by looking at the page, there is a discrepancy between the data in that forecast and the probabilities map.
There isn't a discrepancy. The numbers for "chance of winning" you see when scrolling down are rounded to whole percentages, while the 50.3% figure is rounded to a single decimal point.
I'm not trying to discredit Nate, I'm really not. I'm just trying to work out why myself, Slate.com, and many others looked at his site and saw a very good but not correct prediction and thought "Well, that's safely within the margin of error", and why all the people I've seen on twitter and blogs look at the SAME page and say "Wow, this guy made EPIC predictions and is clearly a witch!"
What do you mean by "very good but not correct"?
His top line prediction, the one right where it says "FiveThirtyEight Forecast: Updated 10:10 AM ET on Nov. 6" and right below " President - Nov. 6 Forecast" says "313.0 electoral votes" in the Obama column.
This is the forecast that I, as a first-time viewer of his site, took to be his final, voting day prediction. Is this so dumb of me? I know for a fact others took the same result, as that was the original one on the Slate.com Pundart graphic.
It's a very good prediction!
Only when scrolling down the page do you see a map of one set of data, that, when compared with the outcome of the state-level voting, matches with another set of data. It can be taken as a forecast of the electoral college votes, but right at the very top of that very same column, the official forecast for the blog is different. And incorrect (pending Florida, Rym).
For some reason, which I can't find by looking at the page, there is a discrepancy between the data in that forecast and the probabilities map.
There isn't a discrepancy. The numbers for "chance of winning" you see when scrolling down are rounded to whole percentages, while the 50.3% figure is rounded to a single decimal point.
Again, that's not clear on the site at all. I saw the map, wanted more information, looked for it, found it, and I'm presented with LESS precise data? And no way to know that that one specific number is rounded to a whole number when every other number above it in the column is rounded to one decimal place? Cheers! I guess it could have been "50.0%", which I could have worked out myself when it wasn't.
Comments
http://fivethirtyeight.blogs.nytimes.com/methodology/
I bring this up because multiple times I've seen posts about Nate Silver, saying he is awesome, and then a link to a dartboard where the closest dart to the center (actually dead center) is another analyst who did the same kind of modeling. If pure accuracy is what you want, Silver isn't the best (and calling two elections in a row is like flipping a coin twice, a quarter of those who do it will get heads twice). If being first is what you want, he wasn't first.
The reason I've enjoyed looking at other similar sites is that they outline their exact process, and invite you to check their methods and numbers and results. I can't find the same invitation to examination on 538.
I think, maybe, that hiding the process, and by adding minor complexity that doesn't noticeably affect the predicted result, Nate Silver seems more mysterious and magical, when really he's not adding all that much.
Who and what exactly are you referring to when you speak of "dead center" or "first"?
The second prong of his popularity is accuracy. And that is not just two presidential elections, but he also has made rather accurate predictions for congressional and gubernatorial races in the U.S. for the last couple of election cycles, plus his history in Baseball predictions.
The third and most recent prong is the blatant antagonism he has drawn from the political right, who tried to discredit him because his prediction is rather trusted. If they succeeded with it they hoped to discredit all projection models which disfavor them by proxy and thus mitigate the "self-fulfilling prophecy" effect that such statistical models can potentially have. Silver stood up to that bullying, and as it shows, he had he was mostly correct in his predictions.
So in sum total, Silver's method may or may not be the best and most accurate prediction model out there, but that isn't the only factor in his popularity or necessarily the reason why people are going "gaga" over him.
I saw a few tweets link to this page, all commenting about how amazing Nate Silver was: Slate's Pundit Scorecard.
The important part of this story isn't that I'm saying Nate Silver isn't accurate, or that his methods aren't clever, but that I was surprised to find, when looking at the Pundarts section, that Nate Silver wasn't the most accurate. Josh Putnam, of FrontloadingHQ, had actually hit the target dead on. And yet, at the time, all these tweeters ignored someone with a more accurate prediction, and were congratulating someone with a less accurate prediction.
The page has now been updated: "Correction, Nov. 7, 2012: ... Nate Silver's prediction had previously been listed as 313 electoral votes. This number was Silver's average of simulated outcomes, not the most likely outcome." I admit that this incorrect reporting of Nate's prediction means he was equally accurate as Josh Putnam (exactly correct), but at the time I saw the tweets, the Nate Silver fans seem to have ignored the fact that someone else was either equally accurate or more accurate, and there was no awareness of that.
Second, First.
Election Projection was set up for the 2004 election. Electoral Vote was set up for the 2004 election. I'm sure there were others, and there have been many more since (like FrontloadingHQ mentioned above).
And when I say "learning form those who went before", the Electoral Vote Votemaster is open about his methods, and in 2004 gave predictions based on three models. From wikipedia:
"The site's final tracking using algorithm 1 posted on Election Day, November 2 gave 262 electoral votes to John Kerry and 261 to George W. Bush, with 15 tossups.[2] The second algorithm (averaging 3 days worth of nonpartisan polls) gave Kerry 245 and Bush 278 with 15 tossups. The third algorithm (predicting the undecideds) predicted 281 for Kerry and 257 for Bush.
The actual vote gave Kerry 252 to Bush's 286. Using nonpartisan polls and averaging a few days worth of polls did best."
If you're going to start four years later than others doing a very similar kind of prediction, this kind of information is pretty handy.
Again, I'm not saying Nate Silver isn't accurate so far, or that his website doesn't look very pretty, or that his graphs are very informative, but that I'm quite bemused about the adulation I'm seeing around the internet. He really isn't a witch. Is he?
Those who just say who wins are making a binary-style statement. A or B. Calling the race one way or the other is something that anyone can do, and anyone with an opinion can simply say it out loud. That's a pundit prediction. And if you are wrong you can bitch about the popular vote, etc.
The other way, to predict the electoral vote, is a more complex beast. It isn't a single binary statement, but the sum of 51 binary statements, each with their own weight.
This means that the Pundart score board is bullshit. The three darts in the center (as of this writing) actually got a score of 51/51.
The first dart out from the center (Donna Brazile) is only "-19" off, which on the face of it makes it look like she is "more accurate" than the next dart (Sam Wang) at "-29". However, this actually means that Donna got 49/51 of the state votes right, and Sam Wang got 50/51.
Everyone who predicted Florida would go to Romney put Obama's electoral vote at 303, not 332, and likewise got a 50/51 result. And this result (Barack Obama: 4,143,362, Mitt Romney: 4,096,346) is crazily close, though admittedly not as close as it has been in the past.
Anyway, what I think Nate Silver is doing differently (though the last forecast I can find on his website is still the 49/51 "Updated 10:10 AM ET on Nov. 6" one) is setting the overall "Chance of Winning" so prominently (90.1%), and tied directly to the electoral vote prediction. Normally the chance of winning is tied to the first kind of prediction, the A or B kind, or to the popular vote. Nate puts it above the popular vote.
However, Nate Silver has presented his result so that the "confidence" rating is waaaaaay broader than the final electoral vote could ever have made the race. EVERY aggregated state-level polling analyst (all the ones I can find) had the same confidence that their model would predict the winner *in this election*, and EVERY ONE was right. This is so obvious that there is even an xkcd about it.
Nate's own information on his own site, in a very clever move, puts the "importance" (probability that a state provides the decisive electoral vote) of Florida's result in 9th place, at just 1.2%. This means that, in Nate's own analysis, calling the correct result in Florida was only 1.2% of his final prediction.
Which, in turn, means that all the other analysts using similar methods, and called Florida incorrectly (understandably when there was fewer than 50,000 in it), were not only 50/51 compared to 51/51 for Silver (though I still can't find his correct predicition on his site), but that the one state they didn't call was only 1.2% significant.
Nate Silver, by having his "uncertainty" so prominently, could easily have got away with calling Florida incorrectly. He states his overall uncertainty above the much smaller number of 1.2%, and this can also overwhelm any difference in the electoral vote.
On the bottom of this page there is a link "Now here's why Nate Silver got it right". And what does that page tell us? Not why, just "that" he got it right. And it's an EPIC Victory for Nate Silver. Really? Epic?
Then again, I'm a big fan of certain bands, and think they have EPIC new albums and EPIC stage shows. I guess, for nerds, Nate Silver is just the star statistician, and I'm the confused onlooker thinking "all he has are pretty graphs and a name".
In short, people are surprised to see that statistics is a thing that actually applies to the real world, and Nate Silver happened to have several factors that put him at the crest of the popularity wave.
That being said, his site is, overall, the best of the bunch; the reasons for this include those already mentioned - good site design and informative graphs. The blog posts are also quite good, and they do in fact give good details on his methodology.
Personally, I value Silver's use of direct and easily accessible probabilistic estimates. "My model assigns an 80% chance to Obama winning Virginia" is much more informative than simply saying they expect Obama to win Virginia, and in particular the clear and accessible probability of overall victory (90.9% chance of Obama victory) is highly convenient.
The only other two predictions that I've seen with this kind of information are these:
http://www.huffingtonpost.com/simon-jackman/pollster-predictions_b_2081013.html
http://election.princeton.edu/2012/11/06/presidential-prediction-2012-final/
On the topic of accuracy, this article seems to be relevant:
http://www.theglobeandmail.com/news/world/us-election/aside-from-nate-silver-which-other-forecasters-and-pollsters-got-it-right/article5034108/
Sam Wang's PEC is OK, and strictly in terms of accuracy the 99.2% chance of Obama victory gets them more points than they lose for getting Florida wrong, but despite their being right 99.2% still seems overconfident to me.
On the other hand, though, it looks like Nate Silver's Senate predictions definitely weren't as good as others, especially his prediction of 92.5% for Republicans to win the North Dakota Senate (which they did not) - as compared to Sam Wang's 75% for Dems to win. Pollster doesn't give probabilities, but it would appear that they, and Sam Wang, correctly predicted all of the Senate seats.
http://fivethirtyeight.blogs.nytimes.com/2012/04/27/arizona-is-probably-not-a-swing-state/
Something that I think needs to be made more clear with regards to which state is the "tipping point" is that it's a matter of correlation, not causation. Despite everything, Ohio is still worth only 16 electoral college votes; Romney winning Ohio would be a big deal only because it would be correlated with Romney victories in other states.
Here's an excerpt "[Silver] considered Florida a virtual toss-up, with Mr. Obama and Mr. Romney tied at 49.8 per cent apiece, but he gave the Democrats a 50.3 per cent chance of taking the state."
That's just an incredibly small number! 0.6% is probably within the margins of error of his own prediction, and this is a prediction of the chance that Obama would win, not even his prediction of the vote count in the state. Which, if it was that close, would have been cause for a recount anyway, which means that Silver's prediction was inside the margin of error for the election itself. That's not good science!
Here's another thing. Nate Silver's page is tricky to understand, and it's difficult to tie down his actual prediction. The top number is the 313 electoral votes for Obama, and this is the one that was chosen by the Slate Pundarts graphic. This prediction was incorrect.
However, scrolling down you find the State by State probabilities map which, if you just go by the shading of the states, is exactly the same as the final state results. However, the shading doesn't show how "strong" the state is for either candidate, like other similar prediction sites (who use poll results to show what the vote count might be), but the likelihood that a candidate will win. This is confusing, and belies the fact that the 50%-60% likelihood that Obama would win Florida way, way broader than, say Electoral-Vote's outlined red shading for Florida, which shows that Romney might get between 1% and 5% more votes than Obama.
What I'm trying to say is that, by not declaring ties, and using 10% ranges for probabilities of winning, rather than much smaller ranges of total state votes, Nate Silver's maps are actually less accurate than other sites. His prediction might have been more correct, but his margin for error was far larger, and only delineated by the 50% cutoff line.
I'm not the only one to be confused by Nate Silver's presentation of his prediction, as the Slate Graphic also was confused about his actual prediction.
The same difficulty of reporting is shown by the linked article in regards to the Electoral-Vote site. They say "The poorest performer out of the more well-known forecasting websites, Electoral-Vote.com, only missed the call in Florida and considered North Carolina a tie (it was won by Mitt Romney). "
The Electoral Vote website didn't think any state would have a tied election. To predict that a state would be exactly split between two candidates would be absolutely bonkers. What the website's map actually showed was that the result was within the margins of error with the polls it was using for its calculation. I think this kind of reporting is actually far more interesting than the "0.3% over 50-50 means I'm calling it for Florida" that Nate Silver's prediction makes. I liked that there were those 15 electoral college votes up for grabs in North Carolina.
The website even says "It is worth emphasizing that the margin of error in most state polls is at least 3% for each candidate... Even some of the ones in the solid light color may technically be statistical ties, but a lead of 5% or more most likely means the candidate is actually ahead."
Finally, Nate Silver has a graph of the probabilities of the actual electoral vote count: "The probability that President Obama receives a given number of Electoral College votes."
332 is at about 20%. 303 is at about 16%. That is a 4% difference between Nate's prediction and the many others who predicted 303. What this shows is that Nate was only 4% more confident in (one of his) actual predictions (the one that you have to infer from the map) over the next most likely (the prediction that many other sites made, and was well within their stated margins of error). His stated number of 313 is hard to find on that graph at all, and so it's unclear how that number is listed at the top of his November 6th prediction.
Lessons learned:
Put at least two different predictions on your website, and when the result comes in, let people pick the one they like best for reporting. If they are a fan, they'll pick the less-accurate but more correct map. If they aren't so sure, they'll pick the top number, and report that you were very close. Either way, you're no worse off than the other sites who were more upfront with their margins of error and less correct.
It really all comes down to Florida, which Nate shows as:
Now-cast 48.2 48.2 Tie
Projected vote share ±2.7 49.8 49.8 Tie
Chance of winning 50% 50% .
All those numbers are the same. He made a call 0.3% "certainty" call for Obama, the popular vote was larger, and he looks like a genius to those who haven't looked closely at his numbers.
Pollster and PEC did give graphs like this, though, so let's have a look at the graphs:
Simon Jackman for HuffPost Pollster:
Sam Wang for Princeton Election Consortium:
(I think the red line is a glitch in the graph here)
Nate Silver for FiveThirtyEight:
Pollster did slightly better here as they assigned ~23% to the actual outcome, and indeed from many other aspects Pollster has indeed been shown to be the most accurate site overall. However, it's ridiculous to criticize FiveThirtyEight for only being 20% confident of an outcome of 332 electoral votes as compared to other sites that don't even assign a confidence number at all.
In simple terms, everything that isn't the darkest shade of blue or red on Nate's map is a case that falls within the model's overall margin for error; the probability numbers indicate the proportion of projected outcomes that give Obama the state vs. giving Romney the state. Considering that the only relevant outcomes are whether Obama or Romney wins a given state, this is a concise representation of the important information at hand.
In effect, Electoral-Vote (which you seem to like) is wasting a lot of resolution on states that aren't really in play. Coloring Minnesota light blue is simply a waste of resolution that's better used on the more interesting states - having the same color scheme for Pennsylvania as you do for Colorado is ridiculous, given that Obama pretty much had Pennsylvania in the bag.
That said, I think the best approach would be to use a continuous color scheme rather than a discrete one. This would in all respects be a better way to present the information at hand, and in such a color scheme Florida would have been white and not red or blue. Clearly no one thinks that Electoral-Vote expected NC to be a tie, but although Obama had decent chances of winning NC based on the available data (as was recognised by FiveThirtyEight and Pollster), it wasn't really a tossup by any reasonable metric.
It appears that Electoral-Vote's choice to shade North Carolina in white was based on a very limited number of polls. By comparison, FiveThirtyEight had North Carolina at 50.6%-48.9% with a 2.6% margin of error, which is indeed something that could go either way but is nonetheless solidly in Romney's favour.
Nate Silver's website does not in any way say "0.3% over 50-50 means I'm calling it for Florida". The probability of ~50% means that the state could very easily go in either direction. It seems to me that your only real complaint is about the color scheme.
All I'm doing, with all of my critique (critique, not criticism), is trying to work out why so many people have gone gaga over Nate Silver compare to other similar sites and analysts.
My point about this map is that, due to his colouring scheme, where a light blue range can cover a predicted outcome 10% wide, and a result that falls within the margin of error (which 0.3% certainly qualifies for) is not distinguished. On another site, this same kind of result is given a white or outlined colour.
What happens is that, when someone looks at Nate's state by state prediction map, and it matches up with the actual results, they are hailing Nate as a genius, DESPITE Nate himself not calling the state of Florida, DESPITE him labeling it a "Tossup", DESPITE him giving polling numbers that show it an exact tie, and DESPITE him predicting an exact 50%-50% chance of it going either way.
And this where I think Nate was really lucky with his chosen algorithm. Even with all of his probabilities, and his top line prediction of 313, and all the rest, it was the way he decided to shade that one map, which compared with similar maps shows one thing (comparative votes) and with his shows something else (probabilities) that made so many people call him a genius.
I'm not trying to discredit Nate, I'm really not. I'm just trying to work out why myself, Slate.com, and many others looked at his site and saw a very good but not correct prediction and thought "Well, that's safely within the margin of error", and why all the people I've seen on twitter and blogs look at the SAME page and say "Wow, this guy made EPIC predictions and is clearly a witch!"
But my broader point stands. Many sites called the election on exactly the same terms. All of them showed that, even with Florida as a tossup, the election wouldn't be decided there. My point is that if, like in 2000, the decision came down to the Florida vote (a 1.9% chance?), Nate Silver was only putting it in the blue by 0.3%. The reason why, up until the election, nobody really cared about this tiny tipping of the scales, well within all margins (of polling and voting), was that Florida just didn't matter so much this time out.
Other states calling put Romney ahead there. It didn't matter for the final result. With the final votes under 50,000, it doesn't matter at all on a national level in terms of popular vote, and it affects nothing in the state itself except as a slight shading on a map. And Nate got that shading right! Along with others.
Florida is bad at this on purpose.
I think that other sites aren't so concerned about forecasting as others though. For example, Electoral Vote (which I'm using an example because I've been following it since 2004, and know the approach the best), while calling itself a vote predictor, is less of a forecasting website and more of a poll tracking website. The map it presents is always of the data within the most recent polls, averaged over a certain number of days. It doesn't take into effect momentum or demographic changes (though those are topics within the blog).
It also doesn't assign weight to different polls and pollsters, and readily admits that bad polling (though declared non-partisan polls) will affect its final score. The votemaster does, of course, acknowledge that certain polls have a bias, and presents a Rasmussen-free results map. And, if you don't like the bare results of it presents, you're free to download the data it uses, every day, and do your own analysis.
Also the votemaster is a non-USA resident libertarian who lives in the Netherlands. Nate obviously has a personality amenable to either demonizing or deifying, depending on the outlook of the journalist involved.
It's down to what you want. I like a more transparent, "this is our data, our method, and our results" approach, and Nate has a "I have secret sauce, and run thousands of simulations, and have all these graphs of uncertainty" approach, which comes off as more magical, or at least more mystical. And when you want to create a modern myth, transparency and sticking with pure data is the opposite of what you need.
The real issue is that people are too obsessed with the idea of "calling" states, and Nate's color scheme doesn't help in this regard. It would be better if it was colored on a continuous spectrum according to the probability, rather than split up into discrete categories - this would have caused Florida to be white on Nate's map, with an undetectable blue tinge.
"State by State Probabilities: Florida Nov. 6th Forecast: 50.3%"
You have to mouse-over the state to see it. Scrolling down, to see more information about Florida, you find all the polling data to be equal: For some reason, which I can't find by looking at the page, there is a discrepancy between the data in that forecast and the probabilities map.
The tipping point rating and percentage has nothing to do with the state by state probabilities, and I'm obviously not being clear in my writing about them. I think, however, this lack of clarity in Nate Silver's presentation is what leads people to see things in his predictions that he himself either isn't stating, or is stating with far less confidence than people are taking away from it.
This is the forecast that I, as a first-time viewer of his site, took to be his final, voting day prediction. Is this so dumb of me? I know for a fact others took the same result, as that was the original one on the Slate.com Pundart graphic.
It's a very good prediction!
Only when scrolling down the page do you see a map of one set of data, that, when compared with the outcome of the state-level voting, matches with another set of data. It can be taken as a forecast of the electoral college votes, but right at the very top of that very same column, the official forecast for the blog is different. And incorrect (pending Florida, Rym).