Who Will Win?
Our forecast, based on a meta-analysis of prediction markets
This wouldn't be a self-respecting blog about forecasting if we didn't come out with our own prediction on Election Eve. Here goes nothing.
But before we dive in, let’s just note how exceptionally close things are. Nate Silver just published his own final forecast:
We ran 80,000 simulations tonight.
Harris won in 40,012.
So ignore any “vibes” you might feel about what will happen tomorrow: the world’s leading political forecaster just estimated the current probabilities of either outcome to be closer to 50-50 than a coin flip.
Now for our forecast. First things first, let’s talk about the methodology behind it, which relied on analyzing prediction markets rather than polling data.
The Raw Signal from Market Makers
When analyzing prediction markets to determine the probability of a Trump victory, the simplest answer might be to look at the explicit market signal: the aggregate price of all betting markets. That’s the service ElectionBettingOdds.com provides, which at time of writing shows Trump at 56.7 percent.
However, that would be overlooking the implicit information revealed by trader behavior. There's reason to believe these markets are skewed by the views of the typically wealthy male demographic that participates in them. Could that imply an overestimation of Trump’s chances?
Let's examine how we might de-bias these prediction markets through the lens of information theory.
Science Interlude: The Kelly Criterion
To understand how we’ve weighted trader positions for our final forecast, let’s walk through a principle from information theory that informed our approach: the Kelly Criterion. This formula helps determine the optimal amount to "bet" on an outcome based on the likelihood of it happening and the payoff if it does.
The Kelly Criterion begins with the reward factor, b. This is the multiple of our "bet" we’d receive if we win. For example, if we wagered $1 and would get $2 back if we won, our b would be 1 (i.e., a 100 percent profit). In the case of a prediction market, an event trading at X percent means we’d win $1 for every X¢ we bet on it, corresponding to a reward factor of 100/X - 1 for every $1 wagered on this market.
Next, the criterion calls for our estimated probabilities — which determine whether or not we think it’s worth buying into this market. Here:
p is the probability of winning, and
q = 1−p is the probability of losing.
The Kelly Criterion leverages both of these to balance expected gains with the risk of losing:
Using the probabilities and our reward factor, we now have an expression that calculates the optimal fraction f of our capital to bet. The overarching idea is to maximize long-term growth by carefully sizing each bet we make.
Top Trader Analysis
In practice, betting the Kelly criterion is considered too high risk for prediction markets, especially as the prices respond very quickly to just a few bets. Most top bettors act closer to what is known as “half-Kelly,” where they bet half of what the Kelly criterion would recommend.
Usefully for us, Polymarket and Manifold Markets both provide full transparency into what the top traders on political markets are betting. This means we can reverse engineer what the most profitable traders think, assuming they are betting at half-Kelly.
A glance through the order books of the main Trump presidency markets1 on either website reveals an interesting pattern: the most profitable traders (by all-time profit and loss) are taking slightly bearish positions relative to the market price: they think his chances are over-priced.
For our forecast, we can simply take a weighted average of all of these top traders’ implied Kelly probabilities, weighting by their total profit on the website (to give more weight to the most successful traders).
Aligned Forecast™
By this method, we calculate a 51.99% probability of Trump winning the presidency. We'd have committed significant Mana to this view on Manifold Markets, but the current probability of 52% already perfectly reflects it.
Of course, this isn’t much more of a confident prediction than Nate Silver’s, although at least we are slightly more certain than a typical coin-flip. Maybe now is a good time to recall the words of poker player Maria Konnikova: “Two percent is a hell of a lot. If I can gain a 2 percent edge over you, I’m on cloud nine.”
That being said, poker players play thousands of hands in a row. Tomorrow is just one election.
Thank you for reading! This blog’s promise to you is to always open-source any analysis code and detailed methodology. However, given the time-sensitive nature of this forecast — it being Election Eve and all — we wanted to publish our prediction first. Full technical details, including the modeling code and data pipeline, will be added here soon.
Manifold: Will Trump win the 2024 Election? Polymarket: Presidential Election Winner 2024