While AI-driven trading hasn’t quite experienced its “iPhone moment”—when everyone has an algorithmic, reinforcement learning portfolio manager readily available—experts suggest that such a breakthrough is on the horizon.
The complexity of trading markets poses significant challenges for AI, as opposed to self-driving cars learning to navigate based on data and algorithms. Predicting market futures remains elusive, despite the vast amount of information available.
This reality complicates the process of refining AI trading models, which have traditionally been evaluated based on profit and loss (P&L). Recent advancements in algorithm customization are leading to AI agents that continuously adapt to balance risk and reward amid varying market conditions.
Integrating risk-adjusted metrics like the Sharpe Ratio into the learning framework significantly enhances testing sophistication, according to Michael Sena, chief marketing officer at Recall Labs. This firm has orchestrated around 20 AI trading competitions, where participants submit their AI agents to compete over four to five days.
“As the next generation of developers searches for alpha in the market, they are focusing on customizing algorithms and accommodating user preferences,” Sena mentioned in a recent interview. “Focusing on an optimized ratio rather than just raw P&L aligns more closely with how leading financial institutions operate in traditional markets. It’s crucial to consider factors like maximum drawdown and value at risk driving P&L outcomes.”
Looking at a more comprehensive view, a recent trading competition on decentralized exchange Hyperliquid, which involved several large language models (LLMs) such as GPT-5, DeepSeek, and Gemini Pro, established a benchmark for AI’s current role in trading. These models were given identical prompts and operated autonomously, but according to Sena, their performance was lackluster, barely exceeding market returns.
“We used the AI models from the Hyperliquid contest and allowed individuals to submit their trading agents to see if they could outperform the foundational models with their specialized approaches,” Sena explained.
The top three positions in Recall’s competition were claimed by tailored models. “While some models underperformed, it became clear that specialized trading agents applying extra logic, inference, and diverse data sources were outperforming the foundational AI,” he noted.
The rise of AI-based trading poses intriguing questions about the future of alpha opportunities, particularly if many are leveraging similar sophisticated machine-learning technologies.
“If everyone utilizes the same agent following the same strategy, does that strategy lose its effectiveness?” Sena pondered. “Is the alpha being detected diminishing because it’s being executed en masse for all users?”
This scenario underlines that those best positioned to take advantage of AI trading advancements will be the ones willing to invest in custom tool development, Sena asserted. As seen in traditional finance, the most effective tools that yield substantial alpha often remain exclusive.
“Many wish to keep these tools as confidential as possible to safeguard their alpha,” Sena said. “They invest significant resources in acquiring these advantages, exemplified by hedge funds procuring unique data sets or family offices developing proprietary algorithms.
“I believe the optimal outcome will feature a portfolio management product where users can provide input on their trading approach. They could specify, ‘This is how I prefer to trade along with my parameters, so let’s implement something analogous but improved.’”
