Got your grok bots going?
Cool.
They’re fun.
And you start to realise how the asset management firm of the future will be run with bots.
But just one thing.
Grokbots, like Claue will soon be a commodity.
Everyone will have them.
The secret sauce then becomes the data layer.
How you build and configure this data layer is the secret sauce and your advantage versus other asset managers.
Here are 8 tips for building an agentic asset manager:
#1 Build a firmwide agent data lake: This is obvious, but needs to be said up front. There needs to be a common data resource which all the AI agents can access. You can’t have everyone playing with bots on their own building everything in silos. In the agentic asset manager AI agents will be updating files, sending notifications and sharing work so there needs to be common layer they interact with. We will talk about roles later.
#2 Build your own Bloomberg: AI agents work better when they have all the data already rather than request data ad hoc from Bloomberg or Factset every time they need to do anything. So start storing lots of data on your side. store historical prices, transcripts, news, financial results, earnings estimates, economic data, whatever. If it’s a data series you use store 20 years of it. Just do it. Storage is cheap so store it all. Essentially you are building your own custom Bloomberg.
You will see why this is useful.
Say you want an agent to alert you to big moves in economic release data. Well, what’s a ‘big move’? Proactive agents need historical context. The Agent needs 20 years of economic history with calculated standard deviations to know what moves are significant. Obviously, when the data release comes out it’s simpler and quicker to have that 20 year history at hand in your data lake rather than pull a full 20 year history from Bloomberg so it can decide whether to alert you.
Don’t be surprised if you are updating over 10,000 data series/month.
#3 Get good at working with API’s: Part of building your agentic data layer is you are going to have to know how to access datasets yourself using API’s. You will accumulate a whole library of API’s for your agents to access data programmatically. Some of these API’s will be free (like FRED and EAI) while others will be data you pay for.
In addition to your free data API’s you will have paid market data API’s like Massive and FMP. It’s also useful to sign up for a news API like Tavily, which has cleaned up news. Otherwise, your searches end up with lots of junk news which doesn’t relate to what you are looking for. News feeds are cool to add to monitoring dashboards.
Don’t slam the API’s or you will get cut off. Be careful when you are building up your database history. Ask Claude to pace out the API requests so you aren’t hammering the AI will lots of requests all at once. It raises red flags on your account. Do a big data backfill over a few hours.
When you are updating data daily try to do the update a little while after the market close when the high frequency trader activity has settled down for the day and everyone has done their end of the day updates. Again, you are going to be doing a lot of data requests and you don’t want to make your data providers angry. Also try to bucket your data requests if possible.
#4 The importance of Registries and Graphs: You want to store all your research and have your agents access it. Totally. You want to be able to say ‘tell me any research my firm has received on Chinese insurance stocks.’ and bam it shows everything. A 360 view. OK. Assuming you had been storing all of the research sent to your firm the next AI plumbing question is what is the universe of Chinese insurance stocks? The AI agent doesn’t automatically know that. And what about a research report on the growth of Chinese private pensions which doesn’t mention any stocks specifically? And what if there is a report which mentions tickers but not company names? Do you want to see those reports too?
There is an art to AI research ingestion, but a key tool is a securities universe. I call mine YWR-Registry. If the AI is going to map research to securities and issuers there needs to be stable universe to map to along with aliases for those securities and issuers. For example, in the US for Apple you would want to have the legal name, ‘Apple Inc.’, the common name ‘Apple’, and the ticker ‘APPL’ linked so a report mentioning ‘APPL’ still links to ‘Apple’.
And I keep saying ‘issuers’ as well as securties because in a situation where there is a podcast and ‘Sequoia’ says they like AI hardware you want to map that insight to ‘Sequoia’ the issuer rather than a specific fund (ie, ‘Sequoia Amazing Opportunities IV’). Then you can query ‘What does Sequoia think about AI hardware?’ and that podcast comes up. Mapping data and research in the debt space also needs to distinguish between issuers and securities.
Maintaining a securities master like YWR-Registry requires a little bit of maintenance work, which is why firms like Goldman Sachs offer access to theirs for a fee if you don’t want to deal with it.
Which brings us to Graphs. YWR-Graph is a work in progress, but the idea is research topics are broad and it can be challenging to map relevant research to stocks. Let’s say Morgan Stanley writes a primer on GPU’s and that report is stored and ingested in my research library. And let’s say that report on GPU technology never mentions any companies specifically. If I search for all research related to ‘Nvidia’ I might want that report to come up, or maybe not. This is where the graph comes in. The graph is a data structure which relates concepts and topics to stocks. Now it is a judgement call if I want to link ‘Nvidia’ to ‘GPU’. Maybe I have linked ‘Nvidia’ to ‘Blackwell’. In which case whether that Morgan Stanley report comes up in my Nvidia search would depend if they mentioned ‘Blackwell’ GPU’s or not. As I said, these Graphs are a work in progress and require judgement. I find them kind of interesting and they seem like they will be an important piece of AI data infrastructure.
#5 Store your decision making context: You want to be able to chat with your AI about difficult decisions. And you want your AI to help you make better decisions going forward.
To do that the AI needs to know why you do what you do, and what you did in the past. It also needs to know the context of your decisions. What was going on at the time? What were the risks you were afraid of? What was the economic outlook? What were the alternatives? This all needs to get stored through time. It’s really important.
‘Hey Stevie..I’m really wondering if I should buy call options in Nvidia… what do you think?’
‘Hi Ted. It’s a sound idea. We’ve had a 3 std price move and valuations are in the 5th percentile while earnings estimates are upward trending. The set up looks good. But I would suggest owning shares instead of call options. In the past when you have made calls like this (United Healthcare and BP) you underestimated the time it took for the calls to work. 78% of the time you get the stock price direction correct, but it take longer than expected for the idea to work and your options expire useless. This time try buying the shares and being more patient.’
‘Thank you Stevie!! I love you!!’
See what I mean. It helps when the AI has a history of your decision making as well as the counterfactual.
It’s also going to be helpful to AI decision making for the AI to know what was happening in the world at the time. The human brain works in stories and timelines.
We all remember everything going on at the top of the internet bubble in 1999, or the crash of the GFC in 2008. We remember the news articles, we remember the emotions. We remember what the experts were predicting back then. We remember how they all got it wrong.
The challenge is that historical context is hard to recreate. The AI sees the historical data, the charts and what happened next, without the context of what everyone was predicting at the time and why what unfolded was a surprise.
But this is essential context.
That’s why I created YWR-Alexandria. It’s kind of weird but I store a data file of stock prices, futures curves, economic forecasts and news items in Google Cloud each day. Kind of a financial newspaper for the day. The idea is to be able to link public context to the internal firm decision context. The AI can understand both what you were deciding internally with the context of what was going on generally in the world.
#6 Your Data Tools
Claude Code: You are going to need to work with Claude Code in the terminal to build everything. Codex works too, but I’m kind of married to Claude for now.
VS Code: It’s helpful to know how to use VS Code, which is basically MS Word for coding. Claude does almost everything, but I find it useful to use VS code sometimes for updating environment variables (and to see all the files).
Google Cloud: Google Cloud works well, especially if you are ingesting research. It’s a one stop shop. You can set up a Google Workspace for your AI agents (Stevie has its own Google Workspace Account), which means the agents can send emails and build spreadsheets on their own. With Google Workspace you can also set up special inboxes and drives to ingest your research. After the data is ingested it can be stored in the Google Cloud. So everything at Google.
AWS: I find the Google UI easier to work with than AWS, but one nice feature of AWS is Route53, which manages web domains. Google doesn't have that. I like to build web based dashboards so it’s nice to be able to manage all the domains and add DNS records in AWS. I mostly use AWS for Route53 at the moment.
Neon: Google and AWS have their own versions of a Postgres database, but I find it easy to work with Neon. A ‘Postgres’ data base is where you will store data which you work with often and want to update a lot. Kind of like spreadsheets for your AI. You will need one.
DuckDB and Parquet Files: For storing millions of historical datapoints, like the whole EIA database, I’ve found storing the data as Parquet files takes up a lot less space and is cheaper. This is for reference data you want to read, not write or change. Store these Parquet files in a Google Storage Bucket. It’s cheap.
When setting this up with Claude make sure you tag when the data was stored. You want to know which version of the data you have, and when you stored it, if the data is changed in the future. I keep all versions of a data series for up to two years.
Railway: Like with Neon both Google and AWS have servers, tons of them, but Railway is easy to use. It’s a few easy clicks to create a new project, link it to Github and add environment variables. With the CLI (command line interface) Claude can control Railway directly.
I use Railway to run daily price updating files, host agents (like Stevie), and host dashboards. You will also use a server like Railway if you want to connect all your internal data to the Claude app through an MCP. I did this for the YWR MCP (Asset Management as a Service (YWR MCP))
Github: You will need a Github account for your firm to store all your code files. The Railway projects will connect to it.
#7 Agent Roles: In Google Cloud, AWS or Neon you are going to need to create different access roles for your agents. Some agents will only read data, while some will read or write data. ‘Roles’ are a big topic when you get into cloud services; especially so when you deal with agents.
Pro tip: I often give an agent its own read/write database in addition to other databases where it has read only access. I find as you are vibe coding and coming up with new ideas it’s useful for Claude to be able to whip up new tables for your agent without any restrictions. Kind of like a notebook for your agent.
#8 Your AI Sellside: Once you have your data layer in place with your securities universe, the API’s, the Google Workspace, and the Railway server, you can start running your own internal AI analyst team which does research, builds forecasts, listens to conference calls and writes investment reports. I’ve been trialing this with YDX where I ask get the AI to analyse 500 stocks at a time. You are welcome to check out.
Just remember it’s a work in progress and 100% AI research.
And not investment recommendation.
Next Steps:
Get building and get ingesting.
Once your data layer is done, link up your agents and bots and unlock the next level.
I’m so hooked I don’t know how to go back.
Oh, and don’t forget to let me know how it goes.
I’m sure you will build something cool.
Have a good weekend!