Jump Trading: Domain-Adapted Agents for Quantitative Research at Scale

Jump Trading processes terabytes of market data daily, searching for predictive signals across thousands of instruments worldwide. Under tight latency constraints, our pipeline spans raw order book events, NLP signal integration, model training, and live execution. The nature of financial data, from irregularity and non-stationarity to microstructure noise, poses unique challenges and opens new directions in HPC, reinforcement learning, AI, and agent research. In this talk, we describe quantitative research workflows applying foundation models and multi-agent systems to market data. We present results from fine-tuning large language models on textual input to produce live trading signals, and from multi-agent systems that combine a broad range of unstructured sources into forecasts consumed by human traders and automated strategies alike. We also discuss hallucination risk management, strict point-in-time correctness, and the evaluation methodology we apply to ensure rigor and reliability in production.