The Rise of AI in Fantasy Football
Recent advancements in artificial intelligence (AI) are set to significantly enhance predictive capabilities in various fields, including fantasy football. As I delved into the world of Premier League fantasy leagues, I decided to experiment by deploying a swarm of AI agents to manage my team. This venture stemmed from the growing influence of AI in our daily lives, prompting me to discover whether a collective of bots could assemble a winning squad.
My initial exploration of AI’s potential in this realm began a few years ago with a tool from the Turing Institute in the UK. This tool made team selections based on historical fantasy data, yet its suggestions often seemed ludicrous, failing to account for crucial real-world factors such as player injuries or coaching changes.
Developing a Comprehensive AI System
To improve my experience, I enlisted the help of Codex, an advanced AI model. I started by providing it with several pertinent research articles, subsequently instructing it to create a system that could coordinate multiple agents across various stages, consult recent news, and execute predictive models.
The result was an impressive setup that could gather and analyse far more data than I could possibly manage. However, I was taken aback by its initial line-up suggestions, which included the baffling idea of benching Erling Haaland. This prompted me to refine my approach further.
I instructed Codex to assemble a team of agents with distinct roles: some to scan news articles and blogs, others to analyse recent match results, and a group dedicated to critically evaluating the recommendations. This restructuring led to a noticeable improvement in the quality of suggestions. An attempt to scale the number of agents beyond 100 yielded minimal benefits and merely inflated my API bill.
The Evolution of Predictive Models
Despite the challenges, the advancements compared to previous attempts were striking. According to Haifeng Xu, a computer scientist at the University of Chicago who examines AI applications in market contexts, contemporary models possess enhanced capabilities to process vast amounts of real-world data and coordinate complex systems for more accurate predictions.
Xu noted that the increasing intelligence of AI in understanding the world is a pivotal reason for its improvement. His fascination with the burgeoning popularity of prediction markets like Kalshi and Polymarket led him and his colleagues to assess how top commercial models performed in various scenarios.
Their experiments involved predicting outcomes such as the NBA championship winner and the individuals with whom President Trump would meet. They quickly realised that these models were performing remarkably well, sometimes surpassing the average human predictor without requiring specific instructions.
New Tools for Evaluating Predictive Accuracy
The research team developed Prophet Arena, a benchmarking tool designed to assess a model’s predictive accuracy, alongside other frameworks aimed at testing AI’s forecasting abilities. Another initiative, Prediction Arena, spearheaded by Harvard University and Arcada Labs, allocated $10,000 to cutting-edge models to wager on prediction markets over 57 days. While the models consistently lost money during the experiment, researchers noticed a gradual improvement over time.
Xu emphasised that a tailored system allows AI to make predictions more effectively, highlighting that agents with specific tasks—like news reading and data collection—seem to yield the best results.
AI vs. Human Expertise in Prediction
Interestingly, AI still falls short when compared to highly skilled experts known as “superforecasters.” Xu explained that while humans lack the capacity to process the same volume of information, their deeper understanding of the real world often leads to superior outcomes.
However, evidence suggests that the gap between human superforecasters and AI is narrowing. The recent Metaculus Cup demonstrated instances where AI models could outperform some of the best human predictors. Additionally, a testing platform named ForecastBench is expected to provide more insights into AI’s ability to match human skills once the results of an ongoing competition are released.
The Future Role of AI in Prediction Markets
These predictive experiments indicate that, much like mathematics, programming, and law, commerce-related sectors may soon experience significant disruption from AI technologies. It is plausible that exceptional human operators will transition into new roles overseeing and coordinating AI agents equipped to forecast with greater acumen than they can achieve alone.
Currently, my AI’s recommendations have not impressed me; my team has yet to outperform the average in recent weeks, and I have no intention of relying on its suggestions for my selections. Nevertheless, it is clear that AI’s capabilities will continue to evolve, and it may soon dominate fantasy leagues with alarming ease—a prospect that could diminish the thrill of winning through personal merit.