CHAPTERS
- 0:18 – 1:19
High-stakes AI in institutional investing: why getting it right matters
Tushara Fernando introduces Man Group and frames the real-world consequences of deploying AI in a regulated investment manager. She emphasizes that AI errors can directly impact pensions and institutional capital, raising the bar for rigor and governance.
- •Man Group manages $200B+ for pensions, sovereign wealth funds, and institutions
- •AI decisions affect real people’s retirement savings and investments
- •Systematic trading is a major area where AI can add value
- •Regulated environment increases the importance of controls and review
- 1:19 – 3:20
Systematic trading explained through the “fantasy football team” signal analogy
She explains what a trading signal is using a fantasy football analogy: ranking choices, selecting winners, and timing changes before the market reprices. The metaphor bridges intuition (form/fixtures) to systematic ranking of securities.
- •Systematic trading = algorithmic decisions across thousands of securities and markets
- •A signal ranks candidates (long vs short) similar to selecting a lineup
- •Timing and cost matter—acting before price moves is key
- •Signals operationalize an underlying “factor” or idea
- 3:20 – 4:22
Backtesting signals: evaluating ideas without knowing the future
Because future performance is unknowable, Man Group relies on historical simulation to test strategies across many regimes. She describes core performance and risk statistics used to judge whether a signal is robust.
- •Backtest over long histories (e.g., 15+ years) to span diverse market regimes
- •Key outputs: annualized return, drawdown, and Sharpe ratio
- •Backtesting provides evidence, not certainty
- •Workflow repeatability is essential for fair comparison between signals
- 4:22 – 5:56
AI-generated signals in production: what AI did (and what humans still do)
Fernando shares that AI has already proposed and helped productionize real trading signals at Man Group, with humans reviewing outputs. She withholds details of the specific strategy as proprietary, shifting focus to the enabling foundations.
- •AI originated ideas, gathered data, ran backtests, wrote proposals, and helped productionize
- •Humans review for sensibility and controls
- •Signals are running with real capital in a regulated firm
- •Talk focuses on the journey and infrastructure rather than the IP of the signal
- 5:56 – 7:27
The iceberg beneath a signal: hidden workflows and why consistency matters
She argues the signal is only the visible tip; the real complexity is in shared data and research workflows. Without standardized processes, teams can reach conflicting conclusions due to inconsistent methods rather than better ideas.
- •Core foundations: data cleaning, price stitching, outlier detection, backtest infrastructure
- •Inconsistent workflows create incomparable results across teams
- •Shared workflows reduce duplicated effort and improve consistency
- •Standardization is critical when selecting among competing signals
- 7:27 – 8:29
Why off-the-shelf Claude isn’t enough: connecting AI to institutional context via skills
Claude is powerful but lacks firm-specific knowledge, data access, and operational capabilities. Man Group’s approach is to expose internal context through skills rather than retraining or fine-tuning, turning institutional knowledge into leverage.
- •Generic models don’t know internal datasets, systems, or how work gets done
- •Man Group’s “superpower” is decades of institutional workflows and technical capabilities
- •Skills act as the connective layer between AI and proprietary context/tools
- •Focus is on access and orchestration, not model fine-tuning
- 8:29 – 9:32
The adoption trap: lots of skills, but built by power users—not process owners
Man Group initially pushed hard on skill adoption via workshops and hackathons, but saw structural issues. Power users optimized for personal convenience, producing fragmented, non-standard skills that didn’t represent organizational workflows.
- •High adoption efforts: workshops, hackathons, show-and-tells, blog support
- •Most skills were created by enthusiastic users rather than workflow owners
- •Resulting skills were local optimizations instead of enterprise solutions
- •Fragmentation prevents agents from relying on consistent, shared capabilities
- 9:32 – 11:36
Expense-report skill incident: a small bug reveals a big governance problem
A humorous expense-report automation illustrates how unreviewed, user-built skills can create organizational pain. A hard-coded cost center caused expense reports to route incorrectly, highlighting accountability and review gaps.
- •Skill automated expense reports from receipt images
- •Hard-coded cost center routed approvals to the wrong approver
- •No review/ownership model meant bugs affected others downstream
- •Example generalizes to serious risks in research/backtesting workflows
- 11:36 – 12:11
Skills governance as the enterprise unlock: from isolated tools to agent-ready foundations
Fernando explains that governance—not just building skills—enables enterprise-scale agentic use cases. The organization needs common, trusted workflows so agents can reliably execute tasks like systematic research and backtesting.
- •Governed skills enable commonality and reuse across teams
- •Agents need stable, well-defined interfaces to workflows
- •Governance transforms skills from productivity hacks into shared infrastructure
- •This shift enabled more complex applications like systematic trading
- 12:11 – 13:11
Building a skills marketplace: visibility, ownership, testing, lifecycle management
Man Group’s solution is a curated marketplace/library where skills are discoverable, tagged, tested, and owned by workflow owners. This introduces the same care and controls as production software development.
- •Central marketplace where every skill is visible and categorized
- •Workflow owners are accountable for skills; review processes are defined
- •Evals/testing, usage tracking, and lifecycle (including retirement) are built in
- •Managed vs community skills balance governance with experimentation
- 13:11 – 14:13
Demo walkthrough (Maya Knowledge): installing plugins/skills and accessing datasets
She introduces Maya Knowledge as the interface to Man Group’s context store and skill collection. The demo shows tailored skill suggestions per business unit and how plugins bundle related skills like data access.
- •Maya Knowledge hosts skills plus a context store for institutional knowledge
- •Skill suggestions are tailored by business unit; ownership is explicit
- •Plugins bundle multiple skills (e.g., a data plugin for datasets)
- •Skills can be installed into Claude individually or via plugins
- 14:13 – 15:15
Signal research demo: using credit card spend to test a trading hypothesis
Using an alternative dataset skill, Claude finds credit card transaction data and explores whether spending patterns relate to stock performance. The demo plots Amazon spend vs returns and then runs a backtest to evaluate predictiveness.
- •Search and retrieve alternative datasets (e.g., US consumer transactions)
- •Plot Amazon monthly spend against stock price/returns
- •Observe seasonal spikes (Black Friday, Christmas) in spend data
- •Backtest compares signal performance vs buy-and-hold
- 15:15 – 16:16
Scaling research: broaden the universe and distribute compute for backtests
To reduce the chance of a single-name fluke, the workflow expands testing to a broader retail universe. Distributed infrastructure runs per-company workers and aggregates results, powered by a small set of foundational skills.
- •Validate generality by testing across multiple retail companies
- •Use distributed compute: one worker per company, results aggregated
- •Case study used four skills end-to-end to create a signal
- •Real research accounts for seasonality, inflation, and broader security sets
- 16:16 – 20:44
Lessons learned and outcomes: context is the moat, skills are production, adoption is people
Fernando summarizes key learnings: organizational context is defensible IP, skills must be treated as production code, and adoption requires training and workflow redesign. She shares adoption metrics and ties governance directly to systematic trading success and future agent swarms.
- •Organizational context/workflows are the moat; expose them rather than reinvent them
- •Treat skills like production code: ownership, review, testing, retirement
- •Adoption isn’t a licensing issue—requires engagement, training, and workflow rethinking
- •Scale achieved: ~750 Claude Code users; 100+ governed skills plus community skills
- •Takeaway: build a “golden path” from AI to context/capabilities to prepare for agentic systems
