AI Wrote a Trading Signal That Is Running Real Money
Hey,
I read a lot of AI case studies. Most are theatre. A demo, a press release, a few quotes from a chief digital officer about the future of work. The substance is thin. The follow up question of "is this in production" usually gets a redirect.
The talk I want to walk you through this week is the exception. Tashara Fernando, head of data and AI at Man Group, stood up at Code with Claude London 2026 and said this.
There are trading signals running right now in production at Man Group, a regulated investment firm running real capital, that were researched, back tested, and proposed by AI. AI came up with the idea. AI got the data. AI ran the back test. AI wrote the strategy proposal. AI productionized the signal. Humans reviewed the output.
Man Group manages over $200 billion. Their clients are pension funds, sovereign wealth funds, and large institutions. The money belongs to teachers in Canada and metal workers in Japan. When she says the stakes are high, she means it literally.
She did not tell us what the signal was. That is their IP. She told us the foundation that made it possible, and the foundation is the part you can actually use.
Three key takeaways
- AI is now writing trading signals at Man Group that touch real capital. Humans review. AI is at the center of the workflow.
- The thing that made this possible was not a better model. It was skills governance. A common marketplace with ownership, evals, tagging, versioning, and a library style review process.
- Adoption is the easy part. The hard part is making sure the actual workflow owners write the skills, not the power users who happen to use those workflows most.
What systematic trading is, briefly
Systematic trading means algorithmic trading across thousands of securities and hundreds of markets. The unit of work is a signal, which is a ranking strategy. You rank stocks by a factor, you go long the top, you short the bottom, and you back test against fifteen or more years of history to see how the strategy performed.
You compute return, drawdown, sharp ratio. Those numbers tell you something about the strategy. They do not tell you the future. The work AI is now doing at Man Group is the upstream half of that loop. Coming up with ideas, getting the data, running the back tests, writing the proposals, productionizing the winners.
The iceberg
Here is the part Fernando emphasized and it is the part most teams underestimate.
The signal is the visible thing. Underneath it is everything that makes it possible. Cleaning the data. Stitching prices when symbols change. Detecting outliers. Running the back tests. Comparing results across teams.
If different teams use different versions of those workflows, they get different answers. One team's back test looks amazing. Another's looks average. You cannot tell whether one team had a better idea or whether they are measuring differently. That kills the entire ability to compare signals, which is the whole game in systematic trading.
Out of the box Claude is a strong general tool. It does not know your data, your systems, or how your team works. The gap is the same for every firm.
"The first thing that we had to do was teach it. Not by retraining it, not by doing fine-tuning, but by giving it access to our data, our capabilities, and our workflows. That's our superpower."
Skills are the connector. You write a skill that captures one of your workflows, attach it to an agent, and now the model knows how your team does that thing.
The story of getting it wrong first
Man Group went all in on adoption. Workshops with Anthropic. Hackathons. Internal blog. Show and tell sessions. Everyone writing skills. The adoption was, in her words, out of this world.
Then the cracks showed up. The skills were being written by power users. Not by process owners. Each skill was a local optimization for one user, not an organizational solution.
She told a story that captures the failure mode cleanly. There was a guy at Man Group who travelled a lot, hated doing expenses, and wrote a skill that ate pictures of receipts and produced expense reports. It worked. He shared it. A few teammates picked it up.
Then the expense approver came around. Why was Claude generating expense reports against his cost center for people in technology, in the people team, in departments he had no relationship with?
The cost center code was hardcoded in the skill. Nobody had reviewed it. The author was not accountable for it. He thought it was funny. Fernando thought it was funny too. But it was symptomatic.
"People were just codifying their ways of doing things. They weren't the organizational ways of doing things. And in many cases, they weren't actually the workflow owner."
In a sales team that is annoying. In systematic trading where signals run real capital, it is a non starter. If the back testing skill was not written by the back testing workflow owner, you cannot trust the signal.
The marketplace
Their solve was a common marketplace. Called My Knowledge. Acts as the firm's context store.
Every skill is visible. Skills are tagged. Skills have evals. Skills are owned by the actual workflow owner. Usage is tracked. Skills are reviewed. Skills have a lifecycle. Skills get retired.
She compared it to a library. Sections for finance, people, research. The library cares for each item. The library has a process for retiring books. The care is what makes the library a library and not a junk drawer.
Plugins group useful skills, like a data plugin that gives access to the firm's data sets. Individual skills can be installed too, like the data set skill for searching alternative data sets.
The credit card demo
The demo Fernando ran was a back test on credit card data. Specific enough to be useful.
Step one. Use the alternative data set skill. Ask Claude what credit card data sets are available. It returns a US consumer transactions data set.
Step two. Plot Amazon's monthly credit card spend against its stock price returns. Blue bars for spend, line for stock. You can see seasonality spikes around Black Friday and Christmas.
Step three. Back test the signal. Is credit card spend predictive? It performs better than buy and hold. A thousand dollars in 2021 would be twenty five hundred today.
Step four. Could be a fluke for Amazon. Run it across a broader universe of retail companies using the firm's distributed compute. Each company runs on an individual worker. Findings collected and compared.
Four skills did most of the work. Each was written by the workflow owner. Each had evals. Each was visible to anyone in the firm. The signal proposed by AI was reviewed by humans. The signal is now in production.
What this generalizes to
If you do not run systematic trading, the specifics do not apply. The pattern does.
Your organizational context is your moat. The frontier labs are not going to solve it for you. It is not on the internet. They do not know your workflows. You already have the context. You have decades of it. The work is on exposing it.
Skills are how decades of institutional knowledge become usable by AI. Treat them like production code, because that is what they will become.
"Plan your approach before you plan the rollout. Who's going to own the skill? Who's going to review it? How does it get retired? How does it get tested? Decide this before shipping the first skill, not after the hundredth like us."
Adoption is a people problem, not a licensing problem. You have to capture organizational context and rethink workflows, not just augment them. That is training and engagement.
The numbers at Man Group say something about scale. Around seventeen or eighteen hundred people in the firm. Seven hundred and fifty using Claude Code. Developers, quants, the people team, finance. Every department is on it. Over a hundred governed skills and at least as many community skills, all in the library, all visible.
The three things to do
First, decide who owns each workflow. Not who uses it most. Who owns it. That is who writes the skill, reviews changes, decides when to retire it.
Second, build the platform before you encourage adoption. Visibility, tags, evals, versioning, ownership, lifecycle. Without that scaffolding you get the expense report story repeated a hundred times across departments.
Third, treat skills like production code. They will run inside AI loops that touch real systems. The hardcoded cost center in a hobby skill is a small mistake. The hardcoded assumption in a back testing skill is a regulatory problem.
Man Group can have AI propose trading signals that run real capital because they built the infrastructure that makes AI output reviewable, comparable, and tied to the right workflow owner. That is the part you build. That is the part nobody can ship for you.
The frontier labs will keep shipping better models. Your moat is the cleaner workflows, the reviewed skills, the library that captures how your team actually works.
Why the word skill is doing real work here
Skills sit in a useful place between three things people normally try to put into prompts. You have raw model capability, which is general. You have system prompts, which are project specific but flat. You have RAG, which is data heavy but not workflow heavy.
A skill captures a workflow. It is the unit of "how we do this thing here." It has inputs, outputs, dependencies, and assumptions. A back testing skill at Man Group probably names the price stitching method, the outlier detection routine, the universe definition, the rebalance frequency, and the metrics to report. That is the work product of fifteen years of arguments inside the firm about what good looks like, codified in a way an AI agent can apply.
Once you see that, you start to see why governance is non negotiable. A skill is a piece of institutional truth. If your truth is wrong, every signal that uses the skill is wrong. If your truth is right but someone copies it into a side skill and modifies one assumption, you now have two competing truths. The library exists to keep there from being two truths.
The CIO conversation this changes
If you sit in a CIO seat or report to one, this story changes the conversation about AI investment. The standard pitch is some version of "we need to evaluate models" or "we need to pick a platform" or "we need to invest in MLOps." All of those are real but they are the easy half.
The harder half is the workflow inventory. Who owns the back testing workflow. Who owns the customer onboarding workflow. Who owns the claim adjudication workflow. Who is allowed to change those workflows. Who reviews changes. How are conflicts resolved. The companies that win this decade are the ones that can answer those questions, because answering them is what makes skills writable.
Man Group's seven hundred and fifty Claude Code users did not appear because the firm picked the right model. They appeared because the firm built a place where institutional knowledge could be exposed, reviewed, and used. The platform earned the adoption by being trustworthy.
If your firm cannot answer who owns a workflow, that is the work to do before you pick a tool.
What I would ask you
If you are inside a firm trying to do this, where are you in the cycle? Are you in the all in adoption phase, watching power users write everything? Have you hit the expense report moment? Are you building the marketplace and trying to make ownership stick?
Reply and tell me. I read every reply, and I am collecting patterns from teams at different stages of this journey.
Marco