Keep pulling the thread on Gary Collier, Tushar Fernando.
Man Group's token consumption has increased 86 times since January.
A Bridgewater paper found that fine-tuning an open-source Qwen model with proprietary data yielded better results for identifying newsworthy financial information than using frontier US models.
For over a year and a half, Man Group has been developing an AI system with multiple agents to automate the quantitative research process, from idea formation to coding backtests and evaluating results.
Man Group is actively trading client assets using approximately 15 to 20 investment models that were ideated and constructed by an AI agent system and subsequently approved by a human investment committee.
Tushar Fernando observes that the duration of a task an agentic workflow can complete autonomously is doubling approximately every 7 months.
At Man Group, employees in finance, operations, and HR are using agentic coding workflows.
Man Group uses AI to synthesize various data types, including podcasts, alternative data, and broker research, to provide meaningful insights to its portfolio managers.
Gary Collier believes the primary bottleneck for Man Group in deploying AI is the organizational change required to implement new capabilities quickly and safely, not a lack of compute or data.
Man Group ingests nearly a terabyte of tick data per day from most exchanges.
Man Group is currently finding that proper data preprocessing, tagging, and structuring provides more value than fine-tuning AI models.
Tushar Fernando asserts that for quantitative research tasks, the quality of underlying data is more important for generating alpha than using the latest frontier AI models.
Man Group has allocated its AI budget to individual business units, allowing them to manage their own token consumption and spending.