How AI agents can reshape arbitrage in prediction markets

Arbitrage opportunities in prediction markets often exist for seconds, giving AI-driven systems a structural advantage over humans.

Prediction markets aggregate human judgment in theory, but some of their consistent trading opportunities may end up captured by systems that move faster than any person can.

Arbitrage opportunities can show up as brief mispricings, from outcomes that temporarily fail to sum up to 100%, to short delays in how quickly markets react to new information.

Rodrigo Coelho, CEO of Edge & Node, said bots are already scanning hundreds of markets per second, a role that increasingly overlaps with more advanced AI-driven agents.

“Capturing those opportunities requires monitoring thousands of markets and executing trades almost instantly, which is why they’re largely dominated by automated systems,” Coelho told Cointelegraph.

That makes prediction markets a natural next step for AI-driven systems built to exploit short-lived pricing gaps without human input.

Bitcoin and crypto prices haven’t been performing well recently, with BitMine’s Tom Lee calling the current sentiment a “mini-crypto winter.” Meanwhile, prediction markets have emerged as venues where users can bet to profit independently of broader economic conditions.

The rise of prediction markets has also seen opportunities such as what Coelho calls “latency arbitrage,” which rely on short windows too narrow for humans to manually target. He told Cointelegraph:

A recent study found that Polymarket exhibits frequent pricing inconsistencies, allowing traders to construct arbitrage positions. These opportunities arise both within individual markets, where probabilities don’t sum to 100%, and across related markets with inconsistent pricing. The researchers estimated that roughly $40 million has been extracted from these inefficiencies.

Prediction markets are still nascent, but their technology has been improving as well. For example, Polymarket recently introduced taker fees to increase trading costs. Outcomes aren’t finalized immediately, making these strategies less reliable and not always profitable.

Aside from arbitrage, AI agents could increasingly take over activity in prediction markets, raising concerns that automated systems may replicate the same behaviors seen from humans. They are trained on human activity, after all. 

Coelho pointed out that large players can influence outcomes by placing sizable bets on one side, and that more advanced agents could exploit similar dynamics at scale.

“If you have a large pool of money and the market is thin, you can bet on one side and sway the market, like we saw in the election when some French guy put in like [$45 million] on Donald Trump winning,” he said.

Polymarket’s open interest was highest around October and early November of 2024, during the US elections, according to Dune Analytics data. Following a sharp initial decline, it has continued to surge in popularity, with politics leading as the most popular topic, followed by sports and crypto.

Related: Federal regulation looms as 11 states go after prediction markets

Pranav Maheshwari, engineer at Edge & Node, said the rapid improvement of AI agents alongside prediction markets makes such risks more urgent and called for guardrails.

“Up until now, AI agents have medium capability and we give them a lot of permissions. With this medium capability, they have already started acting autonomously,” Maheshwari told Cointelegraph.

Trading itself is undergoing a shift, as automation moves from simple execution bots to more advanced, AI-assisted systems capable of identifying and acting on opportunities in real time.

The systems currently used to exploit market inefficiencies remain largely rule-based, but the tools behind them are evolving.

Archie Chaudhury, CEO of LayerLens, said most retail participants are not using AI agents directly, relying instead on chatbot interfaces like ChatGPT or Gemini for research, while more advanced users are beginning to experiment with automation.

“Some of us simply use coding agents such as Claude Code to create automated bots or algorithms for executing trades, while others take it a step further, using autonomous tools such as OpenClaw to enable the automatic execution of trades and other policies,” he told Cointelegraph.

Related: Do Super Bowl ads predict a bubble? Dot-coms, crypto and now AI

As AI literacy among retail traders rises, agents could broaden access to strategies that were previously limited to institutions, according to Chaudhury. However, this does not eliminate competition, and large institutions are already using AI, though not always publicly.

He added that existing large language model architectures are well suited to interpreting structured financial data, which could lower the technical barrier for building trading systems that would have previously required specialized quantitative expertise.

The same dynamics are already visible across crypto markets, where arbitrage increasingly depends on automation rather than human judgment. As these systems evolve, the edge is shifting execution speed. Those leaning on AI and automation have a clear edge over those that don’t.

Magazine: Agent wastes 14 hours of scammers’ time, LLMs ‘poisoned’ by Iran: AI Eye

Source: CoinTelegraph