What are the key takeaways from this Odd Lots episode?

Key insights include: **Data quality over model quality:** For quant research tasks, structured and tagged proprietary data outperforms access to frontier models. Man Group adds plain-English descriptors to datasets like credit card data, builds a unified semantic layer connecting disparate datasets, and invests in institutional knowledge capture — making data architecture the primary driver of alpha, not model selection.; **Agentic signal generation:** Man Group's AI system autonomously generates trading hypotheses by reading academic papers and labeled datasets, writes investment rationale in plain English, builds backtesting code, and runs validation — all before human review. Fifteen to twenty AI-ideated models have already passed human investment committee review and are actively trading client assets.; **Token consumption management via education, not routing:** Rather than building automated query-routing classifiers to assign tasks to cheaper models, Man Group federates token budgets to individual business units and runs internal education programs. This approach caught basic inefficiencies — like agents processing entire Git command outputs as tokens — and produced employee-driven optimization solutions organically.

What did Gary Collier and Tushara Fernando discuss on Odd Lots?

Man Group CTO Gary Collier and Head of Data & AI Tushara Fernando explain how one of the world's largest hedge funds deploys AI across discretionary and systematic investing, covering agentic workflows, data architecture, token budgeting, and the 86x growth in token consumption since January 2025. Key topics include: **Data quality over model quality:** For quant research tasks, structured and tagged proprietary data outperforms access to frontier models. Man Group adds plain-English descriptors to datasets like credit card data, builds a unified semantic layer connecting disparate datasets, and invests in institutional knowledge capture — making data architecture the primary driver of alpha, not model selection.; **Agentic signal generation:** Man Group's AI system autonomously generates trading hypotheses by reading academic papers and labeled datasets, writes investment rationale in plain English, builds backtesting code, and runs validation — all before human review. Fifteen to twenty AI-ideated models have already passed human investment committee review and are actively trading client assets..

How long is this episode of Odd Lots?

This episode is 52 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

Odd Lots

One of the World's Largest Hedge Funds on Its 86x Growth in Token Spending

July 9, 2026

52 min episode · 2 min read

Gary Collier,Tushara Fernando

Episode

52 min

Read time

2 min

Topics

Personal Finance, Relationships, Investing

AI-Generated Summary

Published Jul 9, 2026

Key Takeaways

✓Data quality over model quality: For quant research tasks, structured and tagged proprietary data outperforms access to frontier models. Man Group adds plain-English descriptors to datasets like credit card data, builds a unified semantic layer connecting disparate datasets, and invests in institutional knowledge capture — making data architecture the primary driver of alpha, not model selection.
✓Agentic signal generation: Man Group's AI system autonomously generates trading hypotheses by reading academic papers and labeled datasets, writes investment rationale in plain English, builds backtesting code, and runs validation — all before human review. Fifteen to twenty AI-ideated models have already passed human investment committee review and are actively trading client assets.
✓Token consumption management via education, not routing: Rather than building automated query-routing classifiers to assign tasks to cheaper models, Man Group federates token budgets to individual business units and runs internal education programs. This approach caught basic inefficiencies — like agents processing entire Git command outputs as tokens — and produced employee-driven optimization solutions organically.
✓Agentic task horizon as the key capability benchmark: The METR benchmark, which measures how long an AI agent can autonomously complete tasks humans would perform, doubles every seven months. Agents can now handle tasks requiring sixteen human hours, shifting the relevant skill from prompt-level interaction to end-to-end workflow design and multi-team orchestration across full application builds.
✓Alpha durability through ecosystem depth, not single repositories: AI lowers barriers to dataset analysis, but trading alpha at Man Group derives from the combination of broker relationships, proprietary market data, decades of backtesting infrastructure, and market access — not any single code repository. Some dataset features will become risk factors as they commoditize, but the interconnected system sustains differentiated returns.

What It Covers

Man Group CTO Gary Collier and Head of Data & AI Tushara Fernando explain how one of the world's largest hedge funds deploys AI across discretionary and systematic investing, covering agentic workflows, data architecture, token budgeting, and the 86x growth in token consumption since January 2025.

Key Questions Answered

•Data quality over model quality: For quant research tasks, structured and tagged proprietary data outperforms access to frontier models. Man Group adds plain-English descriptors to datasets like credit card data, builds a unified semantic layer connecting disparate datasets, and invests in institutional knowledge capture — making data architecture the primary driver of alpha, not model selection.
•Agentic signal generation: Man Group's AI system autonomously generates trading hypotheses by reading academic papers and labeled datasets, writes investment rationale in plain English, builds backtesting code, and runs validation — all before human review. Fifteen to twenty AI-ideated models have already passed human investment committee review and are actively trading client assets.
•Token consumption management via education, not routing: Rather than building automated query-routing classifiers to assign tasks to cheaper models, Man Group federates token budgets to individual business units and runs internal education programs. This approach caught basic inefficiencies — like agents processing entire Git command outputs as tokens — and produced employee-driven optimization solutions organically.
•Agentic task horizon as the key capability benchmark: The METR benchmark, which measures how long an AI agent can autonomously complete tasks humans would perform, doubles every seven months. Agents can now handle tasks requiring sixteen human hours, shifting the relevant skill from prompt-level interaction to end-to-end workflow design and multi-team orchestration across full application builds.
•Alpha durability through ecosystem depth, not single repositories: AI lowers barriers to dataset analysis, but trading alpha at Man Group derives from the combination of broker relationships, proprietary market data, decades of backtesting infrastructure, and market access — not any single code repository. Some dataset features will become risk factors as they commoditize, but the interconnected system sustains differentiated returns.

Notable Moment

Man Group's token consumption grew 86 times between January and July 2025 — a figure the firm did not anticipate. Notably, this growth occurred before implementing any automated cost-control routing, meaning the firm is still in an open experimentation phase despite the scale of consumption already reached.

Know someone who'd find this useful?