Election Day Is Here: Why Polls Don’t Tell the Whole Story

I was on Instagram this week and saw something interesting. I came across two back-to-back posts that had nothing to do with each other: one about the Dodgers-Yankees World Series, the other of a Donald Trump meme at McDonalds. It felt like the perfect snapshot of America this last month—baseball, fast food, and, of course, politics.

As the election draws ever closer, and the entire world waits in suspension, our focus has drawn to various election models. Obviously this doesn't come as a surprise. Like most everyone else, social media—traditional media—are filled with seemingly misrepresented sound bites, a sprinkle of Harris-themed Fortnite maps here, and Trump-oriented McDonalds memes there. While amusing, it doesn't really paint a picture of who is going to win on Tuesday. Or does it?

Like some of you out there, quantitative polling is where we turn to. But if there is anything I have learned in the last two election cycles, it's that polling, in its traditional sense, is not very accurate. That may seem obvious, but the reason polls are not very accurate isn't due to small sample sizes that are, in some form or fashion, mathematically enhanced to make large predictions for the entire election. It's because it's almost impossible to gauge whether an election model is 1) statistically better than literal random guessing and 2) better or worse than any other model. At least, that's what Stanford professor Justin Grimmer wrote in his research paper. Now to be fair, I was put on to this paper through a mutual connection to Austin Park, the research director at VoteHub and a Mathematics and Political Science student at Stanford University.

After reading Austin's writeup of his own thoughts, it made me think completely different about election predictions. What if we live in a time where the amount of shares on a meme, and the loser of the World Series, are better predictors of our next President?

First, a quick disclaimer. By day, I’m the Chief Operating Officer here at Atlas, which means I spend most of my time on things like click-through rates, server configurations, and the occasional password reset—admittedly not the most thrilling stuff. But there’s more to my background. I actually have an MBA in finance with a focus on quantitative analytics. Exciting, right? Okay, maybe not for everyone, but it does give me a different perspective when it comes to data and predictions.

Jokes aside, this entire point of this article is not necessarily about polling specifically. It’s about the unreliability and unpredictability of humans and their sentiment— and how that affects polling. This interest, while niche, largely stems from behavioral economics, and more specifically - Dan Ariely's book Predictably Irrational: The Hidden Forces That Shape Our Decisions. I have read this book twice now. Once in business school, and again more recently. And as I looked to apply some theory to human behavior analysis, I didn't have to look far beyond what is absolutely the most prevalent example of humans making an emotionally charged and often irrational decision: the election of a President.

Ariely’s work highlights a fundamental truth: human behavior is often irrational, and yet, it follows certain patterns that can be studied and, to some extent, predicted. This is where traditional polling falls short. Polls assume that voters are rational actors who will answer truthfully and consistently, but as Ariely and countless election cycles have shown us, that’s rarely the case. People’s decisions are influenced by a range of unpredictable factors—emotions, social pressures, misinformation, and even the framing of the questions themselves. So, if we know that humans are predictably irrational, why do we continue to rely on outdated polling methods that treat them as rational? This brings us back to the core of the discussion: how can we better predict election outcomes by acknowledging the complexity of human sentiment? What if we looked beyond traditional polling and started paying attention to these unconventional indicators - the way people react on social platforms, their search patterns, and yes, even those World Series results I mentioned before? These alternative signals might actually reveal more about voter behavior than standard polls ever could. When you dig into this kind of data, you start seeing patterns in how people really think and feel, not just what they say when someone puts them on the spot with a survey question.

I'm convinced that the key to better election forecasting isn't about perfecting our statistical models—it's about accepting and working with the messy reality of how humans actually behave. By blending what we know about human psychology with sophisticated data analysis, we might finally move past the limitations that have made traditional polling so unreliable. Maybe then we'll develop predictions that actually capture what's really going on in voters' minds, rather than just what they're willing to tell a pollster.

The truth is, understanding voter behavior is complex - just like people themselves. And maybe that's exactly why we need to start looking at it differently.

Let's continue.

The Problem with Traditional Polling: A Dinosaur in the Digital Age

Let’s be honest, traditional polling is starting to feel like a relic from another era. The idea that you can just call up a few thousand people, ask them who they’re voting for, and then project that onto an entire nation is not only outdated—it’s flawed. The assumption behind this method is that people will give honest, consistent answers, but as anyone who has followed recent elections knows, that’s rarely the case. Voter sentiment is fluid, and people are more likely than ever to change their minds, keep their true intentions private, or simply avoid engaging with pollsters altogether. But here’s the real kicker: polling doesn’t just fail because people lie or change their minds. It fails because it’s trying to capture human behavior—a notoriously unpredictable variable—using rigid, linear methods. Looking at Austin Park's research, along with several other experts in the field, there's a glaring issue with how we predict elections: polls treat voters like robots who make purely logical choices. But that's not how real life works. People don't sit down with a spreadsheet comparing policy positions before they vote - they're swayed by their feelings, what their friends and family think, and whatever information (true or not) happens to be floating around online. Which brings me to the real question: if we know standard polls are missing the mark this badly, what other tools could we be using to better understand how people will actually vote?

Now let me clarify. Traditional polling is not just asking people questions, getting their answers, and then announcing 'hey world, X will be our next president'. Maybe it was in past history when people were verifiably more honest. But that isn't the case today. Polling today requires contextualization. Any that contextualization is the hard part.

This is where non-traditional data sources come into play. Think about it: people are constantly leaving digital breadcrumbs that reveal far more about their intentions than a simple phone or online survey ever could. Social media posts, search engine queries, even patterns in consumer behavior—all of these are pieces of a much larger puzzle that can help us understand what’s really going on beneath the surface.

OSINT for Contextualization: A New Approach to Polling

Here Open Source Intelligence (OSINT) finds application. OSINT is about leveraging publicly available data to develop a more complex knowledge of voter sentiment, not about only gathering data from news sources or social media. OSINT uses what people are actually saying and doing online in real time, unlike conventional polling, which depends on self-reported data (and all the biases that follow with it. It's a means to monitor public opinion's internet footprint without personally requesting opinions. OSINT, for instance, can examine trends in social media activity—that is, which candidates are referenced most often or how particular issues are trending across several platforms—instead of depending on people to tell you who they support. It can also monitor search engine trends to identify at any one instant the subjects generating public interest. More dynamically and immediately than any conventional poll could provide, these data points present a picture of voter attitude. Here, though, it gets very fascinating: OSINT reveals why people are saying things in addition to what they are saying. OSINT can find the underlying emotions driving public opinion—such as fear, rage, hope, or frustration—by examining the surroundings of social media posts or search inquiries. This type of sentiment analysis enables a better knowledge of voter behavior and goes much beyond the surface-level information that conventional polls gather.

Non-Traditional Data Sources: Beyond Social Media

While OSINT is incredibly powerful, it’s only one part of the puzzle. There are other non-traditional data sources that can provide equally valuable insights into voter sentiment—sources that most pollsters aren’t even thinking about. Consider consumers' behavior. Although at first look it appears disconnected to politics, trends in consumer spending can reasonably predict election results. For example: those who feel financially comfortable are more likely to support the incumbent party since they associate their financial situation with the current administration. Conversely, when consumer confidence declines—that is, when individuals start reducing spending or substituting less luxurious items for needs—this usually indicates that they are unhappy with the current situation and could be seeking a shift in power. Hard data supports this, not only conjecture. Voting behavior in earlier elections has been shown to be rather closely correlated with economic data such as employment market trends or retail sales figures. Because they reflect real conduct rather than self-reported intentions, some researchers believe that these kinds of economic signals are actually more consistent predictors of election results than conventional polling data.

One notable success story comes from MogIA—a predictive analytics platform that used AI-driven OSINT techniques to accurately predict Donald Trump’s victory in 2016 when most traditional polls had him losing by a wide margin. MogIA didn’t rely on asking voters who they supported; instead, it used machine learning algorithms to analyze millions of data points from across the internet—including social media posts, search engine queries, and even YouTube comments—to gauge public sentiment in real time

The result? A prediction model that was far more accurate than anything produced by conventional polling methods. This kind of predictive modeling doesn’t just offer a snapshot of voter sentiment; it uncovers the underlying currents of public opinion that are often invisible to conventional methods. It’s a reminder that in today’s hyper-connected world, the most valuable insights don’t come from asking questions—they come from listening to the digital noise—and doing what class? Contextualize the data.

Spurious Correlation and Unconventional Prediction Models

Predicting election results sometimes involves what statisticians term as spurious correlations—relationships among two variables that seem to be connected but have no causal link. For a classic example, examine the relationship between drowning rates and ice cream sales; both rise in the summer yet one does not cause the other. (if you want to see some more crazy ones, check out more spurious correlations here). Election predictions can benefit from the same idea. For example, even if there is no clear causal relationship, some non-political events—such as sports performances or economic indicators—may seem to match election outcomes. This gets us to Austin Park's "Five Keys to the Presidency." Park's model initially seems to be an interesting variation in spurious correlation. To forecast presidential election results, his model employs five apparently unrelated variables ranging from whether the stadium of the losing World Series team is situated in a former Confederate state to whether the Edmonton Oilers missed the Stanley Cup final.

And yet, despite the randomness of these factors, his model has successfully predicted several elections. This raises an intriguing question: Can such wildly different methodologies reveal something deeper about how we approach predictions? The answer lies in how we interpret these correlations. While Park’s model might not have any direct causal explanation behind it, it highlights an important point: predictions don’t always need direct causality to be useful. In a more quantitative punchy tagline you have maybe heard: correlation is not causation. In fact, many of the most successful prediction models—like MogIA’s use of OSINT—rely on aggregating vast amounts of seemingly unrelated data points and identifying patterns that traditional methods miss. Spurious correlations are usually just ironic or funny and have no indication to an outcome of basically any scenario (unless the matrix is broken). But they do point to an overarching theme. You obviously cannot make the assumption that the vowel count in a candidates name has anything to do with their chances, but you can't explain why it isn't a useful indicator. At least its fun to think its useful. But I digress.

By now you have looked at the chart and came to the same conclusion that I did. Austin originally made this chart before we knew who was going to the World Series. And by looking at it now, knowing the outcome of all the keys, we can assert a conclusion. IF this chart is right, then you can make an uncorrelated prediction on who might be president. The Dodgers are not in an old confederate state, even if you look back in history with them originating in Brooklyn. And the Yankees? Well, they are called the Yankees. So...obviously not a previous confederately based team.

Election Betting Markets: Putting Money Where Your Mouth Is

Let's talk about prediction markets for a moment. While traditional polling asks people who they support, prediction markets ask them to put actual money behind their conviction. This introduces a fascinating dynamic: financial consequence. Regulated platforms like Polymarket and Kalshi have emerged as interesting data points in election forecasting, not because they're perfect predictors, but because they force participants to think twice about their predictions.

The logic is straightforward: people tend to be more careful with their analysis when their own money is at stake. A recent example perfectly illustrates this: two people have a large sized position of $20,000,000 for Trump to win, on PolyMarket. This isn't just notable for the size of the wager - it's significant because it represents a level of conviction that goes beyond simply telling a pollster your preference. When someone risks that kind of capital, they've likely done substantial analysis beyond emotional preference or party loyalty.

Prediction markets also have another advantage over traditional polling: they update in real-time. Unlike polls which provide snapshots of sentiment at specific moments, market prices continuously adjust based on new information, world events, and participant behavior. This creates a dynamic dataset that reflects not just what people say they believe, but what they're willing to risk money on believing. The markets essentially function as a real-time aggregator of insider knowledge, public sentiment, and analytical research - all filtered through the lens of financial risk.

Are prediction markets foolproof? Absolutely not. One of the main reasons Trump is such a heavy favorite on PolyMarket is due to the sheer volume of bets placed. With over $3 billion in trading volume, it's challenging to gauge how much large bets influence the share price. A single, substantial bet on Trump could shift the market price upward—not because his chances of winning have improved, but simply because of the volume moving in one direction. Election betting reflects public opinion in monetary terms, making it more of a lagging indicator of sentiment rather than a real-time predictor of changing odds.

Then there's the issue of herd mentality in betting markets. A significant factor behind Trump's rising odds was a massive $20 million bet placed on him. This type of wager can lead retail traders to follow suit, thinking, "If someone bet that much, they must know something I don’t." This creates a feedback loop where people pile on without any real change in underlying probabilities.

So, do election betting markets provide a more nuanced view than traditional polling? Almost certainly not. But are they entirely useless for broader analysis? No, they still offer some insights into collective sentiment. They do provide another data point that, when combined with OSINT and other non-traditional indicators, helps build a more complete picture of likely election outcomes. After all, asking someone who they think will win is one thing - asking them to bet their savings on that prediction is quite another.

A Multi-Faceted Approach: The Future of Election Forecasting

So where does this leave us? If there’s one thing we’ve learned from recent elections (and from behavioral economics), it’s that human behavior is far too complex to be captured by any single method—especially one as limited as traditional polling. The future of election forecasting lies in adopting a multi-faceted approach that combines traditional polls with non-traditional data sources like OSINT, search engine trends, consumer behavior patterns, and even the World Series. By integrating these diverse datasets into a cohesive model—and using advanced statistical techniques like machine learning to analyze them—we can create more accurate and reliable forecasts that reflect the true complexity of voter sentiment.

In short, while spurious correlations like Austin Park’s model may seem like statistical quirks, they remind us that predictions don’t always need to follow conventional logic. By embracing diverse methodologies—from sports outcomes to AI-driven OSINT—we can build more robust models that reflect the complexity of modern elections. After all, in an era where voter behavior is increasingly unpredictable, sometimes the most unconventional approaches can yield the most accurate results.

In the end, forecasting elections might be less about pinning down certainties and more about exploring patterns in the unpredictable. Traditional polling has its value, but our digital lives offer new, nuanced data that can reveal unexpected layers of voter sentiment—from online trends and economic indicators to cultural touchpoints like sports and memes. These tools, while unorthodox, remind us that election outcomes are influenced by a mix of logic, emotion, and context, often in surprising ways. So as you get out to vote tomorrow, if you haven't already, it’s worth remembering that no model or OSINT methodology can capture the whole story. At the end of the day, it’s not predictions that shape outcomes; it’s the people who show up and cast their votes.