An outline of how we are approaching the looming era of LLM-driven applied research.
For a biased-but-useful preview of the systematic strategy that we are launching in the first half of 2023, visit this page. Six funds/prop shops are receiving our signal in real time. For information on Dr. Judge and Prof. Ratkovic, you can go here.
An artificially intelligent “agent” is a term for a large language model (LLM) that is hooked up to other software systems and given discretion when to employ those systems. The LLM, with initial human direction, decides how to write code and conduct research, at times requesting human feedback. LLM-based agents are only months old, though the topic is growing rapidly in the open source community. It is a commonly held opinion that the design of agents is a more open, important question than even that of further advances in the underlying LLMs themselves.
Research - quantitative and qualitative,
Below, a case is made for a business focused on building agents that sit on banks and funds’ client-side servers but are capable of navigating a purpose-built data center. The data center stores petabytes of real-time market data and a large number of GPUs, which the agents can use to run AI and machine learning tasks, returning models and data to the client side.
OpenAI’s release of ChatGPT kicked off a race to connect LLMs to software systems and the rest of the internet. These tools are beginning to mature. When that process is complete, the most important wave of technological change since the creation of the internet will begin.
With these tools, one can connect ChatGPT to a firm’s private information; allow it to search the internet for events that occurred seconds ago; instruct it to write, run and test code; and create “agents” that break complex problems down into small steps, spin up tens of different AIs - including versions of itself - to address sub-problems, and then iterativaly test the whole solution chain.
Funds that do not intelligently integrate these systems will quickly become antiquated.
It may sound reductive, but as software begins to write software, there are so many feedback loops it becomes difficult to make predictions. We can, however, sketch a few parts of the picture.
Example: What are the five most-anticipated data releases coming out of Seoul next week? Build me a calendar and dashboard that presents consensus estimates, the estimates of our fund’s economists, and a real-time chart of the number of social media posts about each release. As the results are released, juxtapose those numbers with the average number of posts from the past two years.
Example: Get me the order book data for the top 20 commodity futures from the CME’s API. How far back does the data go? OK, use GANs to generate a distribution of how my strategy would have done in the past three days.
Example: Here is a screenshot of my Bloomberg launchpad. Rebuild it using data providers and open source JavaScript and Python modules. Spend no more than $100 dollars a month on API keys to do this.
Example: Create a thirty parge report on my performance during the past year. Under what conditions did I do better or worse than other traders on my desk? What were the major news stories during the worst 10% of my portfolio’s days? Rank their frequency by category.