Building An Investment Analyst

Despite the hype, real-world applications of AI are limited (thus far) — The best way to learn AI is to use it — I am building an investment analyst.

VIEW AS PDF

BUILDING AN INVESTMENT ANALYST

I have built an investment AI to support my research efforts. Using only my desktop computer and open-source models I now have a virtual “analyst.”

Despite the hype surrounding AI, finding real-life applications for the technology remains challenging. Large language models (LLMs) often hallucinate, refuse to reveal their “logic,” and struggle to interact with the world. They can be difficult, even contemptuous, employees.

Despite these limitations, I have found several productive uses for my analyst. But first, where did my analyst come from?

Building AI

It is surprisingly simple to operate an LLM on a local PC. By following a series of YouTube videos, I downloaded Meta’s open-source Llama3 model. This model takes up roughly 20GB of memory and, when queried from my PC’s terminal, provides answers without requiring an internet connection. It is truly “my AI.”

My basic Llama analyst encodes substantial information from the internet and is capable of providing some low-level support. The performance is sluggish though, so after selling off several GPUs from my crypto mining phase, it looks like I’ll be back in the market.

The Written Word

As an LLM, RefAI (my analyst) is exceptional at proofreading. If you’ve noticed an improvement in these writings, the credit belongs to RefAI. Once I write a paper, I simply provide it to RefAI for helpful rewrites and wording changes.

While RefAI is well-versed in value investing concepts, it does not yet understand how to create unique content. Asking for an entire paper, or even an idea for a white paper, results in boring, well-trodden concepts familiar to most investors. This is similar to asking a fresh college graduate to write on an investment topic.

Data Retrieval

RefAI also excels as a data retrieval agent. RefAI is at my side while conducting investment analysis and saves time by quickly providing macroeconomic information. This is crucial for error checking. For example, I may compare the GDP of Japan with that of the APAC region when sense-checking a market size estimate.

The best use of RefAI in this context has been finding average annual exchange rates. I often benchmark across numerous global currencies, and RefAI has helped me quickly standardize my models into US dollars. Previously, this required downloading daily or monthly FX rates and computing annual averages. Notably, RefAI does not actually compute but instead has annual FX rates encoded in its knowledge base.

Giving RefAI a RAG

RAG, or retrieval-augmented generation, is the technical term for providing an AI with specific data to analyze. Most functional uses of AI today rely on the RAG technique. An example would be providing your proprietary HR manual to an AI to answer company-specific questions.

With my RAG setup, RefAI has unique attributes that I can call upon. I provided RefAI with over 100 transcripts from IT distribution companies across the globe. I then instructed RefAI that it was “trained in CIA language interpretation techniques.” I asked RefAI to evaluate these transcripts for tone, honesty, and hidden risks.

Scaled asset managers can hire CIA interrogators to provide training on management interviews and transcript analysis. Today, RefAI can provide a similar service to me—for free.

Here is my CIA analyst in action: 

Putting a Face on a Name

I can also ask RefAI for prompts to be provided to other AI agents. Below is RefAI’s preferred image of “itself.”

Apparently RefAI, like most introverted investment analysts, is a little shy, preferring to generate images that are more representation than physical detail.

DeepSeek & Next Steps

The next step will be upgrading RefAI’s brain to one of the new DeepSeek models. While a larger-parameter model will be helpful, the real value will be implementing DeepSeek’s “Reasoning Engine.”

This reasoning engine will, at a high level, allow RefAI to iterate, producing results but then asking itself if the answer appears correct. Through multiple iterations the ultimate output should be far more valuable, with the added benefit of seeing RefAI’s “logic.”

More to come in the next few weeks!