FiveThirtyEight

One unique aspect of the site is Silver's efforts to rank pollsters by accuracy, weight their polls accordingly, and then supplement those polls with his own electoral projections based on demographics and prior voting patterns. "I did think there was room for a more sophisticated way of handling these things", Silver said.[12][24]

FiveThirtyEight weighs pollsters' historical track records through a complex methodology[538 8] and assigns them values to indicate "Pollster-Introduced Error".

At base Silver's method is similar to other analysts' approaches to taking advantage of the multiple polls that are conducted within each state: he averaged the polling results. But especially in the early months of the election season polling in many states is sparse and episodic. The "average" of polls over an extended period (perhaps several weeks) would not reveal the true state of voter preferences at the present time, nor provide an accurate forecast of the future. One approach to this problem was followed by Pollster.com: if enough polls were available, it computed a locally weighted moving average or LOESS.

However, while adopting such an approach in his own analysis, Silver reasoned that there was additional information available in polls from "similar" states that might help to fill the gaps in information about the trends in a given state. Accordingly, he adapted an approach that he had previously used in his baseball forecasting: using nearest neighbor analysis he first identified "most similar states" and then factored into his electoral projections for a given state the polling information from "similar states". He carried this approach one step further by also factoring national polling trends into the estimates for a given state. Thus, his projections were not simply based on the polling trends in a given state.

Furthermore, a basic intuition that Silver drew from his analysis of the 2008 Democratic party primary elections was that the voting history of a state or Congressional district provided clues to current voting. This is what allowed him to beat all the pollsters in his forecasts in the Democratic primaries in North Carolina and Indiana, for example.[11] Using such information allowed Silver to come up with estimates of the vote preferences even in states for which there were few if any polls. For his general election projections for each state, in addition to relying on the available polls in a given state and "similar states", Silver estimated a "538 regression" using historical voting information along with demographic characteristics of the states to create an estimate that he treated as a separate poll (equivalent to the actually available polls from that state). This approach helped to stabilize his projections, because if there were few if any polls in a given state, the state forecast was largely determined by the 538 regression estimate.

In July 2008, the site began to report regular updates of projections of 2008 U.S. Senate races. Special procedures were developed relying on both polls and demographic analysis. The projections were updated on a weekly basis.[538 9]

In the final update of his presidential forecast model at midday of November 4, 2008, Silver projected a popular vote victory by 6.1 percentage points for Barack Obama and electoral vote totals of 349 (based on a probabilistic projection) or 353 (based on fixed projections of each state).[538 10] Obama won with 365 electoral college votes. Silver's predictions matched the actual results everywhere except in Indiana and the 2nd congressional district of Nebraska, which awards an electoral vote separately from the rest of the state. His projected national popular vote differential was below the actual figure of 7.2 points.

The forecasts for the Senate proved to be correct for every race, but the near stalemate in Minnesota led to a recount that was settled only on June 30, 2009. In Alaska, after a protracted counting of ballots, on November 19 Republican incumbent Ted Stevens conceded the seat to Democrat Mark Begich, an outcome that Silver had forecast on election day.[25] In Georgia, a run-off election on December 2 led to the re-election of Republican Saxby Chambliss, a result that was also consistent with Silver's original projection.

During the first two months after the election, no major innovations in content were introduced. A substantial percentage of the articles focused on Senatorial races: the runoff in Georgia, won by Saxby Chambliss; recounts of votes in Alaska (won by Mark Begich), and Minnesota (Al Franken vs. Norm Coleman); and the appointments of Senatorial replacements in Colorado, New York, and Illinois.

After President Obama's inauguration, Sean Quinn reported that he was moving to Washington, D.C., to continue political writing from that locale.[538 11] On February 4, 2009, he became the first blogger to join the White House press corps.[538 12] After that time, however, he contributed only a handful of articles to FiveThirtyEight.

During the post-2008 election period Silver devoted attention to developing some tools for the analysis of forthcoming 2010 Congressional elections,[538 13][538 14] as well as discussing policy issues and the policy agenda for the Obama administration, especially economic policies.[538 15][538 16] He developed a list of 2010 Senate races in which he made monthly updates of predicted party turnover.[538 17]

Later, Silver adapted his methods to address a variety of issues of the day, including health care reform, climate change, unemployment, and popular support for same-sex marriage.[26] He wrote a series of columns investigating the credibility of polls by Georgia-based firm Strategic Vision, LLC. According to Silver's analysis, Strategic Vision's data displayed statistical anomalies that were inconsistent with random polling. Later, he uncovered indirect evidence that Strategic Vision may have gone as far as to fabricate the results of a citizenship survey taken by Oklahoma high school students, which led him to denounce Strategic Vision as "disreputable and fraudulent".[27][538 18][538 19][538 20][538 21][538 22][28][lower-alpha 1] FiveThirtyEight devoted more than a dozen articles to the Iranian presidential election in June 2009, assessing of the quality of the vote counting. International affairs columnist Renard Sexton began the series with an analysis of polling leading up to the election;[538 23] then posts by Silver, Andrew Gelman and Sexton analyzed the reported returns and political implications.[538 24]

FiveThirtyEight covered the November 3, 2009, elections in the United States in detail.[538 25][538 26] FiveThirtyEight writers Schaller, Gelman, and Silver also gave extensive coverage to the January 19, 2010 Massachusetts special election to the U.S. Senate. The "538 model" once again aggregated the disparate polls to correctly predict that the Republican Scott Brown would win.[538 27]

In spring of 2010, FiveThirtyEight turned a focus on the United Kingdom general election scheduled for May 6, with a series of more than forty articles on the subject that culminated in projections of the number of seats that the three major parties were expected to win.[538 28] Following a number of preview posts in January[538 29] and February,[538 30] Renard Sexton examined subjects such as the UK polling industry[538 31][538 32][538 33] and the 'surge' of the third-party Liberal Democrats,[538 34] while Silver, Sexton and Dan Berman[lower-alpha 2] developed a seat projection model. The UK election was the first time the FiveThirtyEight team did an election night 'liveblog' of a non-U.S. election.[538 35]

In April 2010, The Guardian published Silver's predictions for the 2010 United Kingdom General Election. The majority of polling organisations in the UK use the concept of uniform swing to predict the outcome of elections. However, by applying his own methodology, Silver produced very different results, which suggested that a Conservative victory might have been the most likely outcome.[30] After a series of articles, including critiques and responses to other electoral analysts, his "final projection" was published on the eve of the election.[538 36] In the end, Silver's projections were off the mark, particularly compared with those of some other organizations, and Silver wrote a post mortem on his blog.[538 37] Silver examined the pitfalls of the forecasting process,[538 37] while Sexton discussed the final government agreement between the Conservatives and the Liberal Democrats.[538 38]

On June 6, 2010, FiveThirtyEight posted pollster rankings that updated and elaborated Silver's efforts from the 2008 election. Silver expanded the database to more than 4,700 election polls and developed a model for rating the polls that was more sophisticated than his original rankings.[538 39][31]

Silver responded on 538: "Where's the transparency? Well, it's here [citing his June 6 article], in an article that contains 4,807 words and 18 footnotes. Every detail of how the pollster ratings are calculated is explained. It's also here [referring to another article], in the form of Pollster Scorecards, a feature which we'll continue to roll out over the coming weeks for each of the major polling firms, and which will explain in some detail how we arrive at the particular rating that we did for each one".[538 40]

As for why the complete 538 polling database had not been released publicly, Silver responded: "The principal reason is because I don't know that I'm legally entitled to do so. The polling database was compiled from approximately eight or ten distinct data sources, which were disclosed in a comment which I posted shortly after the pollster ratings were released, and which are detailed again at the end of this article. These include some subscription services, and others from websites that are direct competitors of this one. Although polls contained in these databases are ultimately a matter of the public record and clearly we feel as though we have every right to use them for research purposes, I don't know what rights we might have to re-publish their data in full".

Silver also commented on the fact that the 538 ratings had contributed to Markos Moulitsas's decision to end Daily Kos's use of Research 2000 as its pollster.[32]

Subsequently, on June 11, Mark Blumenthal also commented on the question of transparency in an article in the National Journal titled "Transparency In Rating: Nate Silver's Impressive Ranking Of Pollsters' Accuracy Is Less Impressive In Making Clear What Data Is Used".[33] He noted that in the case of Research 2000 there were some discrepancies between what Silver reported and what the pollster itself reported. Other researchers questioned aspects of the methodology.[34]

On June 16, 2010, Silver announced on his blog that he is willing to give all pollsters who he had included in his rating a list of their polls that he had in his archive, along with the key information that he used (poll marginals, sample size, dates of administration); and he encouraged the pollsters to examine the lists and the results to compare them with the pollster's own record and make corrections.[538 41]

In September 2014, Silver put into the public domain all of his pollster ratings,[538 42] as well as descriptive summary data for all of the more than 6,600 polls in his data collection for the final three weeks of U.S. Presidential primaries and general elections, state governor elections, and U.S. Senate and U.S. Congress elections for the years 1998–2012.[35] In addition to updating his pollster ratings, he published an updated methodological report.[538 43]

When the transition to The New York Times was announced, Silver listed his staff of writers for the first time.[40] However, of the seven listed writers, only three of them had published on 538/New York Times by late December 2010: Silver, Renard Sexton and Hale Stewart. Andrew Gelman contributed again in early 2011.[538 46] Brian McCabe published his first article in January 2011.[538 47][lower-alpha 3]

Beginning in 2011, one writer who emerged as a regular contributor was Micah Cohen. Cohen provided a periodic "Reads and Reactions" column in which he summarized Silver's articles for the previous couple of weeks, as well as reactions to them in the media and other blogs, and suggested some additional readings related to the subject of Silver's columns. Silver identified Cohen as "my news assistant".[538 48] Cohen also contributed additional columns on occasion.[538 49]

On September 12, 2011, Silver introduced another writer: "FiveThirtyEight extends a hearty welcome to John Sides, a political scientist at George Washington University, who will be writing a series of posts for this site over the next month. Mr. Sides is also the founder of the blog The Monkey Cage, which was named the 2010 Blog of the Year by The Week magazine".[538 50]

While politics and elections remained the main focus of FiveThirtyEight, the blog also sometimes addressed sports, including the March Madness[538 51][538 52][42] and the 2012 NCAA Men's Basketball tournament selection process,[538 53] the B.C.S. rankings in NCAA college football,[43] the NBA,[538 54][538 55][538 56] and Major League Baseball matters ranging from the 2011 attendance at the New York Mets' Citi Field[538 57] to the historic 2011 collapse of the Boston Red Sox.[538 58]

In addition, FiveThirtyEight sometimes turned its attention to other topics, such as the economics of blogging,[538 59] the financial ratings by Standard & Poors,[538 60] economists' tendency to underpredict unemployment levels,[538 61] and the economic impact and media coverage of Hurricane Irene (2011).[538 62][538 63]

Adapted from a FiveThirtyEight October 2011 graph published in The New York Times.[538 64]

FiveThirtyEight published a graph showing different growth curves of the news stories covering Tea Party and Occupy Wall Street protests. Silver pointed out that conflicts with the police caused the sharpest increases in news coverage of the protests[538 64] and assessed the geography of the protests by analyzing news reports of the size and location of events across the United States.[538 65]

FiveThirtyEight rolled out its 2012 general election forecasting model on June 7, 2012. The model forecast both the popular vote and the Electoral College vote, with the latter being central to the exercise and involving a forecast of each state. In the initial forecast, Barack Obama was estimated to have a 61.8% chance of winning the electoral vote. The website provided maps and statistics about the electoral outcomes in each state as well as nationally. Later posts addressed methodological issues such as the "house effects" of different pollsters as well as the validity of telephone surveys that did not call cell phones.[538 66]

On the morning of November 6, Election Day, Silver's model gave President Obama a 90.9% chance of winning a majority of the electoral votes.[538 67] The 538 model correctly predicted the winner of all 50 states and the District of Columbia.[44][lower-alpha 4] Silver, along with at least two academic-based analysts who aggregated polls from multiple pollsters, thus not only correctly predicted all 50 states, but also all nine "swing states".[45] In contrast, individual pollsters were less successful. For example, Rasmussen Reports "missed on six of its nine swing-state polls".[46]

An independent analysis of Silver's state-by-state projections, assessing whether the percentages of votes that the candidates actually received fell within the "margin of error" of Silver's forecasts, found that "48 out of 50 states actually fell within his margin of error, giving him a success rate of 96%. And assuming that his projected margin of error figures represent 95 percent confidence intervals, which it is likely they did, Silver performed just about exactly as well as he would expect to over 50 trials. Wizard, indeed".[47][48] Additional tests of the accuracy of the electoral vote predictions were published by other researchers.[49][50]

On September 3, 2014, FiveThirtyEight introduced its forecasts for each of the 36 U.S. Senate elections being contested that year.[538 69] At that time, the Republican Party was given a 64 percent chance of holding a majority of the seats in the Senate after the election. However, Silver also remarked, "An equally important theme is the high degree of uncertainty around that outcome. A large number of states remain competitive, and Democrats could easily retain the Senate".[538 70] About two weeks later, the forecast showed the Republican chances of holding the majority down to 55 percent.[538 71]

FiveThirtyEight sought to apply its mathematical models to the Oscars, and produced internal predictions regarding the subject, predicting four out of six categories correctly.[53] The website also compiled a list of other predictions made by other people using different methods.[54]

FiveThirtyEight applied two separate models to forecast the 2016 presidential primary elections – polls-only and polls-plus models. The polls-only model relied only on polls from a particular state, while the polls-plus model was based on state polls, national polls and endorsements. For each contest, FiveThirtyEight produced probability distributions and average expected vote shares according to both models.[55]

As early as June 2015, FiveThirtyEight argued that Donald Trump "isn't a real candidate".[56] When Donald Trump became the presumptive Republican nominee in May 2016, New York Times media columnist Jim Rutenberg wrote that "predictions can have consequences" and criticized FiveThirtyEight for underestimating Trump's chances. He argued that by giving "Mr. Trump a 2 percent chance at the nomination despite strong polls in his favor ... they also arguably sapped the journalistic will to scour his record as aggressively as those of his supposedly more serious rivals".[57]

In a long retrospective, "How I Acted Like a Pundit and Screwed up on Donald Trump", published in May 2016 after Trump had become the likely nominee, Silver reviewed how he had erred in evaluating Trump's chances early in the primary campaign. Silver wrote, "The big mistake is a curious one for a website that focuses on statistics. Unlike virtually every other forecast we publish at FiveThirtyEight – including the primary and caucus projections I just mentioned – our early estimates of Trump's chances weren't based on a statistical model. Instead, they were what we [call] 'subjective odds' – which is to say, educated guesses. In other words, we were basically acting like pundits, but attaching numbers to our estimates. And we succumbed to some of the same biases that pundits often suffer, such as not changing our minds quickly enough in the face of new evidence. Without a model as a fortification, we found ourselves rambling around the countryside like all the other pundit-barbarians, randomly setting fire to things".[58]

On the Democratic side, FiveThirtyEight argued that Senator Bernie Sanders could "lose everywhere else after Iowa and New Hampshire"[59] and that the "Democratic establishment would rush in to squash" him if he does not.[60] Sanders went on to win 23 states in the primaries.

Fairness and Accuracy in Reporting, a progressive nonprofit media watch group, wrote in May 2016 that FiveThirtyEight "sacrificed its integrity to go after Sanders" and that they have "at times gone beyond the realm of punditry into the realm of hackery – that is, not just treating their own opinions as though they were objective data, but spinning the data so that it conforms to their opinions."[61]

FiveThirtyEight's predictions for each state primary, both for the Republican and the Democratic party nominations, were based on statistical analysis, not on the analyst's opinions. The core data employed were polls, which FiveThirtyEight aggregated for each state (while also considering national polls) using essentially the same method it had employed since 2008. In the 2016 primaries, the projections also took into account endorsements.[62] The website also kept track of the accumulation of national party convention delegates.[63] In a comparison of prediction success published by Bloomberg News after the primary season was completed, FiveThirtyEight's prediction success tied for the highest percentage of correct primary poll winners, at 92%; but it lagged behind PredictWise in predicting a larger set of primaries.[64] Notably, even with FiveThirtyEight's track record of correctly predicting elections that pollsters get wrong, it still missed Bernie Sanders's upset victory in the Michigan primary, for instance, regarded as "one of the biggest upsets in modern political history".[61]

The final prediction by FiveThirtyEight on the morning of election day (November 8, 2016) had Hillary Clinton with a 71% chance to win the 2016 United States presidential election,[65] while other major forecasters had predicted Clinton to win with at least an 85% to 99% probability.[66][67] FiveThirtyEight's model pointed to the possibility of an Electoral College-popular vote split widening in final weeks based on both Clinton's small lead in general polls, but also on Trump's improvement in swing states like Florida or Pennsylvania, mixed with Clinton's poor performing in several of those swing states in comparison with Obama's performance in 2012.[68] The main issues pointed out by the forecast model was the imbalance of Clinton's improvement in very populated states like Texas, Georgia (projected safe for Republican) and California (projected safe for Democrats);[68] mixed with her inability to attract white voters without a college degree, an increasing demographic in swing states, in addition to a potential decline in turnout from minorities.[69] In consequence, Clinton's probabilities to win the Electoral College were not improving.[68] Silver also focused on state-by-state numbers in so-called 'must-win' states like Ohio and Florida, plus a consideration of polls' margin of error in advantages of less than three points.[70]

Donald Trump won the election. FiveThirtyEight projected a much higher probability of Donald Trump winning the presidency than other pollsters,[66] a projection which was criticized by Ryan Grim of the Huffington Post as "unskewing" too much in favor of Trump.[71] While FiveThirtyEight expressed that "nonetheless, Clinton is probably going to win, and she could win by a big margin", the forecaster also made a point about the uncertainty of poll trackers in some cases, about a considerable number of undecided voters and about the unpredictable outcome in traditional swing states.[72]

In early August 2020, FiveThirtyEight announced that for their 2020 general election forecast they had designed a new graphical structure. This included going with modular structure, a "ball swarm" design for the chart depicting each candidate's chances, and the addition of a "forecast mascot" named Fivey Fox.[538 72] An episode of "Chart Chat" discussing the design described the direction saying "FiveThirtyEight has leaned heavily towards a cutesy and engaging approach. The Fivey Fox mascot pops up next to most charts with call-outs to more further information."[73]

Fivey Fox would also issue reminders to readers of "the potential for extreme outcomes" according to Jessica Hullman, in a piece written for The Hill. Hullman also said of the design that the introduction of the mascot, in conjunction with the new simplified look of the page, was "perhaps the strongest indicator that Silver intends to emphasize uncertainty" in his coverage of the 2020 election.[74]

Jasmine Mithani, visual journalist with FiveThirtyEight, said in an interview when asked about the complaints of new mascot said "I think the biggest complaint about Fivey Fox is that some people find it infantilizing, but that wasn't our intention" and that the motivation for including the character was to help make the forecast more of "a teaching tool".[75] In November 2020, Rolling Stone reported that Fivey Fox had reached over 7,000 followers on his dedicated Twitter account.[76]

The forecast favored the actual winner of 48 states, the District of Columbia, and four of the five congressional districts awarding electoral votes, only missing Florida, North Carolina, and Maine's 2nd congressional district. In those three contests the forecast had favored Biden, but they were carried by Trump. Despite correctly forecasting Biden to win nationally, they overestimated Biden's margins in some battleground states such as Wisconsin, Michigan and Pennsylvania and underestimated Trump's margins in states such as Ohio, Iowa and Texas. Their forecast showed Democrats winning Senate races in North Carolina and Maine, which Republicans ended up winning. However, their forecast did correctly predict that the Democrats would take control of the Senate. In the House elections, their forecast favored Democrats to gain seats, yet Democrats suffered a net loss of 10 seats. The actual house results fell outside their 80% confidence interval, with Democrats winning 222 seats, lower than the confidence interval's lower bound of 225.[77]