"So, what does your AI say about who will win the election?" One use case of the US election these days is to deride GenAI. Can GenAI outperform the pollsters and pundits? Everyone knows the answer—of course not, as there are more GenAI models than pollsters, and each model at any point can provide more different answers than voters on the ground. Still, the question is essential in analyzing the predictive powers of these models. Knowing the limitations is necessary when the answers from these models could appear far more assertive than any predictor in real life if prompted to forsake the diffidence.
In Coin Tosses to Elections, AIs Can Add to the Noise with Assurity
Let’s start with something deceptively simple: a coin toss. Picture a coin being flipped into the air — heads or tails, a fifty-fifty outcome. Now imagine trying to predict which side will land up. Theoretically, if we had access to every tiny bit of information about that toss—the precise force used, the angle, air resistance, gravitational effects, even the physical properties of the coin—we could predict the result. These inputs are impossible; hence, no amount of processing of a small set of approximate data is likely to yield perfect predictions.
With some data, pre-GenAI statistical and computational tools probably changed the odds in some instances where data quality was good. With GenAI equations in the fray and the ability to analyze non-quantitative data, one can conceptualize scenarios where they, too, may increase the odds, but only in specific cases with massive data stacks and still with results not materially better than any other pre-GenAI models.
Despite their apparent differences, predicting election outcomes shares certain similarities with predicting a coin toss. While the data surrounding elections is far more complex, it remains, in essence, a chaotic system filled with numerous variables: voter sentiment, last-minute swings, turnout rates, external crises, and more. These factors are constantly shifting, and even the most intelligent AI cannot fully grasp them to conclude what will happen collectively when voters vote.
Generative AI models can scan mountains of voter data, social media sentiment, and economic indicators. But they are still working with incomplete information. The algorithms can estimate and forecast based on historical trends. Still, elections are not wholly predictable—they are often driven by individual human behavior, which can shift dramatically in response to the right stimulus. Even the most sophisticated AI, armed with terabytes of data and intricate algorithms, can only provide a probabilistic estimate akin to a highly advanced opinion poll.
GenAI: Can Make One a Market Guru, but Not a Market Oracle
Investor pitches emphatically tout GenAI's use in portfolio construction these days, implicitly assuring better returns. That begs the question of whether GenAI can genuinely enhance investment returns. If the answer isn’t a clear 'yes,' then skeptics might rightfully ask: if GenAI tools can’t guarantee improved results, why should they overhaul their current investment strategies?
The answer, unfortunately, is highly nuanced. AI is undoubtedly transforming how we analyze information, fundamentally changing how we conduct research. This transformation parallels the impact of calculators and spreadsheets when they first entered the financial analyst’s toolkit. Imagine processing vast datasets with a slide rule—what was once our primary analytical weapon now seems almost unimaginable. Every new technology in the last few decades kept reshaping our ability to extract knowledge out of ever-growing sets of data, and GenAI is the next leap forward. We've documented these changes before and can share specific examples from our own workflows to illustrate how GenAI is reshaping financial analysis.
Analyzing is Different from Predictions: A Somewhat Technical Section on the Evolution of Quantitative Methods in Financial Markets
Despite these advancements, the idea that AI can reliably predict stock prices—whether for the next day, month, or longer—is unrealistic, akin to predicting a coin toss or election outcome. While a few advanced tools might slightly improve their success rates by incorporating GenAI parameters and methods, for most investors, it merely adds complexity and more variables to manage.
To simplify, imagine 10 stocks. Theoretically, this results in 1,024 possible combinations of indices or baskets based on which stocks are included or excluded. The theoretical number of combinations goes to infinity once the weights of any kind are included. Using linear quantitative methods, as early as in the late 1990s, the diligent could data mine hundreds of viable portfolio construction methods by trying out countless combinations to extract those that would pass all the backtest results of any stock or instrument universe the testers selected.
By the 2000s, one could make the equations more dynamic and environment-dependent (like a car's navigation system that reroutes based on traffic conditions), with the non-linearity and ticker-price-based signals added soon after. The theoretical combinations one could evaluate likely climbed to Cantor's higher levels of infinity, and the sophisticated investors moved away from relying on data-mining-based results and more to what logically made sense, along with more stringent risk management tools.
The complexity of what one can do in such models explodes with the inclusion of new AI methods, tools, and possible quantifications or parameterizations of all things subjective. Theoretically, the number of possible portfolios that appear to outperform in backtests could exceed the grains of sand on Earth, even approaching the estimated number of atoms in the observable universe. However, most of these are likely mirages, overfitting past data rather than capturing authentic market patterns. Cynically, the most significant use case of the backtest-driven models has been in attracting the non-quantitatively oriented into black box investment methods that allowed its champions to forsake logic and analysis for otherwise inscrutable decision generators. To be clear, these are not the same as highly disciplined risk-based models of the best in the space who rarely try to impress solely through backtest methods.
Just like we upgraded from log tables to calculators to spreadsheets, AI is the next quantum jump, adding new dimensions to our decision-making power and also, the levels of complexities. But here's the tool-invariant kicker: in financial markets, the average stays average. No matter how fancy our tools are, not everyone can win. Some might ditch AI and still do fine, like those who stick with darts instead of spreadsheets. But the pros will adapt, embracing AI for deeper analysis and smarter efficiency. It's not about guaranteeing everyone wins but about raising the bar for the whole game.
Predictions: A Lamentation of AI's Limitations?
If AI can’t predict elections or guarantee investment outperformance, where does it shine? We have answered this numerous times in previous articles. In short, its power lies in its ability to work with relationships and interconnections—identifying and analyzing the interplay between variables in a way all previous tools simply cannot. This is true across almost all domains we can think of at present. GenAI based equations can show us areas to investigate for better chances of success - irrespective of the domain - but it will not be able to assure outcomes, almost ever.
Which brings us back to elections and the ability of AI to predict them—or not. Perhaps the most valuable lesson from AI's attempt to predict elections or markets is this: in trying to predict the unpredictable, we've created the world's most sophisticated way to remind ourselves of our limitations. It's as if we've built a supercomputer just to tell us what ancient philosophers knew all along – the future remains stubbornly uncertain, even with terabytes of data and quantum computing power at our disposal.
The real victory isn't in perfect prediction, but in better understanding the present and improving our decision-making processes.
