Every four years, as the U.S. Presidential election draws near, data enthusiasts eagerly dive into the world of forecasts and predictions, while the broader public often faces a mix of excitement and anxiety. With the election date of November 5th rapidly approaching, everyone is asking the same question: Who is most likely to win? And there is no shortage of forecasters ready to provide their answers.
The Common Ground Among Election Forecasters
At first glance, many election models seem to rely on similar data: state-level polls, national polls, economic indicators, and approval ratings. While these factors offer valuable insights, the U.S. Presidential election ultimately hinges on winning a majority of the 538 electoral votes allocated among the 50 states. As a result, most forecasters focus more heavily on state-level polling, using simulations to estimate each candidate's probability of winning in individual states, which are then aggregated into a national forecast. This is particularly the case as the election draws nearer, with the broader signals receding into the background. We are not yet at that stage, however.
Key Decisions in Polling Models
As election day nears, forecasters face two critical decisions:
Aggregating State-Level Data: Combining state-level polls and other relevant data to estimate the likelihood of each candidate winning the electoral votes assigned to each state.
Formulating a National Forecast: Deciding how to aggregate these state-level probabilities into a single national outcome, especially when accounting for the potential for correlated polling errors across states.
The weight given to these decisions can vary depending on how close the race is. In a tightly contested race, small adjustments in state-level probabilities can significantly impact the overall forecast. Conversely, in a race where one candidate has a clear lead, a particular challenge is to accurately factor in the possibility of correlated errors, which can heavily influence the probability assigned to less likely outcomes.
Diverse Approaches to Election Forecasting
Currently, several major models are providing live forecasts for the 2024 election, each with its unique methodology:
FiveThirtyEight: A well-known model now under new leadership, which has recently refined its methodology and actively adjusts polling inputs based on new data.
The Economist: Applies an evolved version of the Votamatic model, which has shifted to a more closed-source approach but still provides a general overview of its methods.
Princeton Election Consortium (PEC): Focuses heavily on state-level polling but has been criticized for not adequately addressing the correlation of polling errors across states, an issue that significantly affected its 2016 predictions.
PollyVote: Founded in 2004, PollyVote emphasises combining multiple forecasting methods, such as polls, prediction markets, expert judgment, and econometric models, to enhance accuracy. Over time, it has added new components, such as citizen forecasts, continuously refining its approach to align with evidence-based principles.
Decision Desk HQ: This model employs a range of machine-learning techniques to predict election outcomes. Although the precise details of its approach are not fully transparent, it ultimately uses a simulation-based forecast similar to other models. Decision Desk HQ is known for its fast reporting on election night, combining its sophisticated modelling with a robust data-gathering network.
Data Diary: Data Diary adapts The Economist’s 2020 model into a more fully Bayesian framework, allowing it to handle uncertainty and new information more dynamically. The model uses complex statistical methods to integrate multiple data sources, giving it the flexibility to adjust to shifts in public opinion and other election dynamics.
Three other notable platforms that provide unique insights into election outcomes are Betfair, Polymarket, and PredictIt:
Betfair: One of the largest and most established form of prediction market, Betfair allows participants to bet on a wide range of political outcomes. As a major player in the online betting world, Betfair aggregates the collective sentiment of users on numerous markets, including U.S. elections. The market odds fluctuate in real-time based on the volume and direction of bets, providing a “wisdom of the crowd” perspective that reflects public sentiment and market confidence in a candidate's chances.
Polymarket: A blockchain-based prediction market that also offers betting on political events, Polymarket enables users to buy and sell shares in various election outcomes using cryptocurrency as the medium of exchange. Because it is decentralised and leverages blockchain technology, Polymarket tends to attract a tech-savvy audience that may offer different insights from traditional prediction markets. The market's odds shift based on trading activity, providing a dynamic view of public opinion.
PredictIt: An online prediction market launched in 2014, PredictIt allows participants to buy and sell shares on the outcome of political events, using a continuous double auction format. It is very influential, with over 160 academic data-sharing partners.
Current Predictions and Points of Disagreement
Most models agree on a some key points, for example tha: Kamala Harris has improved upon Joe Biden’s polling numbers, and that a very small number of key states are crucial to the overall outcome. However, they diverge on who currently holds the lead. FiveThirtyEight, The Economist, PollyVote, Decision Desk HQ, and the Princeton Election Consortium, currently lean towards Harris, the latter two albeit very marginally. The Silver Bulletin and Data Diary give Trump the edge, albeit only Nate Silver’s Bulletin by more than a whisker. Meanwhile, prediction markets like Betfair and PredictIt currently tilt to Harris, in line with major bookmakers, while Polymarket also leans to Harris, but only by the finest of margins. The much-heralded presidential debate between Harris and Trump, which the Vice President is widely acknowledged to have won decisively, is only just beginning to filter into the polls. It will be fascinating to see what immediate difference, if any, that makes.
Looking Ahead
In the coming weeks, it will be interesting to examine each of these models more closely, assessing their strengths and weaknesses, their historical accuracy, and the reasons behind their different predictions. Until the actual results come in after November 5th, though, that is all they are – forecasts – and some will prove rather more prescient than others. I have my own thoughts on that, but I’ll just say that I tend, based on my own published research, to trust at any point in time the market aggregate, if I’m pushed to choose, over the outlier.