How Jump Trading is scaling quant research with ChatGPT

Jump Trading, a quantitative trading firm, has integrated GPT‑6 Astra to tackle longer, more ambiguous research challenges by automating predictive modeling workflows. Lucas Baker, Head of LLM R&D, notes that even slight improvements in prediction accuracy over random chance can yield profitable strategies at scale. The firm uses market data, news, events, and alternative sources to build models, but markets' complexity often makes precise forecasting impossible. GPT‑6 Astra enables agents to handle tasks ranging from coding to validating new hypotheses, transforming AI from a tool for minor tasks into a system capable of developing entire codebases independently.

Baker describes AI as most effective when treated like a colleague, with researchers defining problems, environments, and evaluation criteria while steering agents in real time. Agents now autonomously analyze findings, redirect efforts based on predefined standards, and stack incremental improvements without constant human oversight. This recursive process allows long-running tasks to pull from multiple data sources, prioritize key insights, and interrelate findings into comprehensive analyses. Baker emphasizes that the system can operate for days, making subtle judgments about relevance and significance without manual intervention at each step.

The firm operates across diverse asset classes and time horizons in a heavily regulated industry where mistakes carry financial and compliance risks. Baker highlights the importance of balancing AI-driven innovation with robust safeguards, including system design, clear constraints, and human review. AI is applied not only to feature development but also to quality control, security, and monitoring. Outputs like trading signals undergo rigorous review, treated as potentially informative but fallible, and integrated into controlled execution environments to mitigate risks.

Baker envisions a future where "autoresearch"—recursive improvement of systems by agent researchers—becomes standard in quantitative workflows. Today, even advanced setups require human check-ins to define tasks, evaluate intermediate results, and adjust priorities. Future systems may rely on loosely structured agent fleets coordinated by other agents, autonomously exploring ideas, allocating compute, and integrating findings from open-ended questions. Baker reflects on rapid advancements, noting that models progressed from writing single files in 2024 to creating entire codebases in 2025 and tackling open research questions in 2026.