Skip to content
ClaudeClaude

Trading signals that trade themselves

There are trading signals in production at Man Group right now — running real capital — that were researched, backtested, and proposed by AI. Tushara Fernando shares what made that possible inside a regulated investment firm: a governed skills framework and core data layer that taught Claude Code how Man Group's quants have worked for decades, now scaled across ~750 developers and quants and 100+ skills. Walk away with the governance model that lets compliance say yes to AI on your most load-bearing workflows.

Tushara Fernandoguest
May 21, 202620mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Man Group scales AI trading signals through governed skills marketplace foundation

  1. Man Group uses AI to assist systematic trading by generating, backtesting, writing up, and productionizing trading signals, with humans reviewing outputs due to high financial stakes.
  2. The core blocker to scaling AI wasn’t model capability but workflow inconsistency—different teams’ data-cleaning and backtesting pipelines produced incomparable results.
  3. Early “skills” adoption failed because power users built local optimizations without process-owner accountability, creating errors and governance issues (illustrated by a hard-coded expense cost-center incident).
  4. They shifted to a governed skills marketplace where workflows are owned, reviewed, tested with evals, tracked, and managed through a lifecycle—turning skills into reusable organizational primitives.
  5. With this foundation, Claude can compose governed skills (datasets, plotting, distributed compute, backtesting) to explore alternative data like credit-card spend and evaluate signal performance across a broader stock universe.

IDEAS WORTH REMEMBERING

5 ideas

In systematic trading, shared workflows matter as much as ideas.

A signal is the “tip of the iceberg”; inconsistent data stitching, outlier handling, and backtest setups can make one team’s results look better than another’s for non-scientific reasons.

Treat AI skills like production code from day one.

Skills quickly become operational dependencies for agents and humans, so they need testing (evals), review, monitoring, and a retirement path—planning this after mass rollout creates debt.

Power-user-built automations don’t scale without process ownership.

Local optimizations can encode incorrect assumptions (e.g., hard-coded expense cost center), so the workflow owner—not just an enthusiast—must be accountable for correctness and maintenance.

Governance is the enabling layer for agentic automation.

A visible, tagged marketplace with consistent, vetted skills gives agents reliable tools to act; without commonality and standards, agents can’t safely chain actions across the enterprise.

Organizational context is a durable moat in AI adoption.

Frontier models are general-purpose; competitive advantage comes from exposing proprietary workflows, datasets, and infrastructure to the model through controlled interfaces rather than retraining.

WORDS WORTH SAVING

5 quotes

We manage real people's money, thousands of people's pensions and investment capital. So when we think about AI, the stakes are high for us.

Tushara Fernando

There are trading signals running right now in production at Man Group, a regulated investment firm running real capital that were researched, backtested, and proposed by AI.

Tushara Fernando

So coming up with the signal is the quick bit. The hard part is everything that you need, everything that's underneath it, all of the workflows that make it happen, that allow you to act on it.

Tushara Fernando

Adoption is not a licensing problem. It's a people problem.

Tushara Fernando

Focus on that organizational context. That is your IP. It's your moat. It's one of the few safe spaces left in AI.

Tushara Fernando

Systematic trading signals and backtesting metrics (drawdown, Sharpe)Workflow standardization and comparability across teamsClaude “skills” as connectors to internal data and capabilitiesGovernance: ownership, review, evals, lifecycle, visibilityMarketplace/library model: managed vs community skills, pluginsAlternative datasets (e.g., credit card transactions) for alpha researchEnterprise adoption: training, engagement, and workflow redesign

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.