CHAPTERS
Why AI in asset management has high stakes
Tushara Fernando frames Man Group’s responsibility: AI decisions impact real pensions and institutional capital. She sets the context for why reliability, control, and accountability matter when applying AI to investing.
Systematic trading explained through the “fantasy football team” analogy
She demystifies trading signals by comparing them to selecting players for a fantasy team. Signals rank a universe of choices and allocate “back” versus “avoid/short” based on a strategy.
Backtesting: validating signals across market regimes
Because the future is unknown, Man Group relies on historical simulation to evaluate whether a signal is robust. She describes key evaluation metrics and why consistent measurement is essential.
AI-generated signals in production (and why the details stay proprietary)
Fernando reveals that AI has already originated, tested, documented, and helped productionize real signals at Man Group—with human review. She clarifies she won’t disclose the specific alpha ideas, focusing instead on the enablement journey.
The iceberg beneath the signal: shared workflows as the real differentiator
She emphasizes that the visible “signal” is just the tip; the hard work is standardized data and research workflows. Without shared workflows, teams get inconsistent results and comparisons become meaningless.
Teaching Claude your organization without fine-tuning: skills as the connective tissue
Claude is powerful but generic; it doesn’t know Man Group’s data, systems, or ways of working. Skills provide controlled access to internal context and capabilities so AI can operate inside real enterprise workflows.
Early adoption trap: power users built local optimizations, not enterprise workflows
Their first push maximized skill creation via workshops, hackathons, and show-and-tells, but governance was missing. Power users encoded personal shortcuts that didn’t generalize and weren’t owned by process accountable leaders.
The expense-report incident: a small bug that exposed a big governance gap
A humorous-but-serious example illustrates how unguided skills can cause operational problems. A hard-coded cost center routed expense approvals to the wrong person, revealing lack of review and ownership.
Skills governance as the secret sauce: marketplace, ownership, testing, lifecycle
Man Group shifted to a governed model where skills are curated like a library and treated as production assets. This creates dependable building blocks that agents can safely chain for complex tasks like trading research.
Demo walkthrough: the “Maya Knowledge” skills library and context store
She introduces Man Group’s internal platform where skills and context are organized by business unit and governance level. Users can install individual skills or bundled plugins to quickly access approved datasets and tools.
Building a signal with alternative data: credit card spend vs. stock returns
Using skills, they explore a dataset of US consumer transactions and compare spend patterns to Amazon’s stock performance. They then backtest whether spend peaks predict returns and compare results to buy-and-hold.
Scaling research across a universe using distributed compute skills
To avoid single-name flukes, they expand from Amazon to a broader retail universe. Skills orchestrate distributed workers per company and aggregate results, demonstrating how standardized tooling enables scalable experimentation.
Key lessons and operating principles for enterprise AI
Fernando summarizes what they learned: context is the moat, skills become production code, and adoption requires people/process change. Planning governance early prevents chaos later.
Outcomes at Man Group and the path to “swarms of agents”
She closes with adoption metrics and the strategic vision: governed skills enable AI to operate safely at scale. The end-state is many agents leveraging trusted workflows to continuously search for investment opportunities.
