I saw some finance bro in Twitter rant about how Grok was the best thing since sliced bread. This LLM, developed by xAi, has built-in web search and reasoning capabilities… and people are losing their shit at what they perceive it can do for financial analysis tasks.
Like yes, this is better than GPT, which doesn’t have access to real-time information, but you can build a MUCH better financial assistant in about an hour.
And yes, not only is it extremely easy, but it it also works with ANY LLM. Here’s how you can build your own assistant for any task that requires real-time data.
If you know anything at all about large language models, you know that they don’t have access to real-time information.
That is, until Grok 3.
You see, unlike DeepSeek which is boasting an inexpensive architecture, Elon Musk decided that bigger is still better, and spent over $3 billion on 200,000 NVIDIA supercomputers (H100s).
He was leaving no stone left unturned.
The end result is a large language model that is superior to every other model. It boasts a 1 million token context window. AND it has access to the web in the form of Twitter.
However, people are exaggerating some of its capabilities far too much, especially for tasks that require real-time information, like finance.
While Grok 3 can do basic searches, you can build a MUCH better (and cheaper) LLM with real-time access to financial data.
It’s super easy.
Even language models like Grok are unable to perform complex analysis.
Complex analysis requires precise data. If I wanted a list of AI stocks that increased their free cash flow every quarter for the past 4 quarters, I need a precise way to look at the past 4 quarters and come up with an answer.
Searching the web just outright isn’t enough.
However, with a little bit of work, we can build a language model-agnostic financial super-genius that gives accurate, fact-based answers based on data.
Doing this is 3 EASY steps:
Retrieving financial data for every US stock and uploading the data to BigQuery
Building an LLM wrapper to query for the data
Format the results of the query to the LLM
Let’s go into detail for each step.
Using a high-quality fundamental data provider like EODHD, we can query for accurate, real-time financial information within seconds.
We do this by calling the historical data endpoint. This gives us all of the historical data for a particular stock, including earnings estimates, revenue, net income, and more.
Note, that the quality of the data matters tremendously. Sources like EODHD are the perfect balance between cost effectiveness and accuracy. If we use shit-tier data, we can’t be surprised when our LLM gives us shit-tier responses.
Now, there is a bit of work to clean and combine the data into a BigQuery suitable format. In particular, because the volume of data that EODHD provides, we have to do some filtering.
Fortunately, I’ve already done all of the work and released it open-source for free!
We just have to run the script
And the script will automatically run for every stock and upload their financial data.
Now, there is some setup involved. You need to create a Google cloud account and enable BigQuery (assuming we want to benefit from the fast reads that BigQuery provides). But the setup process like this is like any other website. It’ll take a couple minutes, at max.
After we have the data uploaded, we can process to step 2.
This is the step that makes our LLM better than Grok or any other model for financial analysis.
Instead of searching the web for results, we’ll use the LLM to search for the data in our database. With this, we can get exactly the info we want. We can find info on specific stocks or even find novel stock opportunities.
Here’s how.
Step 1) Create an account on Requesty
Requesty allows you to change between different LLM providers without having to create 10 different accounts. This includes the best models for financial analysis, including Gemini Flash 2 and OpenAI o3-mini.
Once we create a Requesty account, we have to create a system prompt.
Step 2) Create an initial LLM prompt
Our next step is to create a system prompt. This gives our model enough context to answer our questions and helps guide its response.
A good system prompt will:
Have all of the necessary context to answer financial questions (such as the schemas and table names)
Have a list of constraints (for example, we might cap the maximum output to 50 companies)
Have a list of examples the model can follow
After we create an initial prompt, we can run it to see the results.
Then, we can iteratively improve the prompt by running it, seeing the response, and making modifications.
Step 3) Iterate and improve on the prompt
Once we have an initial prompt, we can iterate on it and improve it by testing on a wide array of questions. Some questions the model should be able to answer include:
What stocks have the highest net income?
What stocks have increased their grossProfit every quarter for the past 4 quarters?
What is MSFT, AAPL, GOOGL, and Meta’s average revenue for the past 5 years?
After each question, we’ll execute the query that the model generates and see the response. If it doesn’t look right, we’ll inspect it, iterate on it, and add more examples to steer its output.
Once we’ve perfected our prompt, we’re ready to glue everything together for an easy-to-read, human-readable response!
Finally, once we have a working system that can query for financial data, we can build an LLM super-intelligent agent that incorporates it!
To do this, we’ll simply forward the results from the LLM into another request that formats it.
As I mentioned, this process is not hard, is more accurate than LLMs like Grok, and is very inexpensive. If you care about searching through financial datasets in seconds, you can save yourself an hour of work by working off of what I open-sourced.
Or, you can use NexusTrade, and do all of this and more right now!
NexusTrade – a free, UI-based alternative for financial analysis and algorithmic trading
NexusTrade is built on top of this AI technology, but can do a lot more than this script. It’s filled with features that makes financial analysis and algorithmic trading easy for retail investors.
For example, instead of asking basic financial analysis questions, you can ask something like the following:
Additionally, you can use the AI to quickly test algorithmic trading strategies.
Finally, if you need ideas on how to get started, the AI can quickly point you to successful strategies to get inspiration from. You can say:
You can also browse a public library of profitable portfolios even without using the AI. If you’d rather focus on the insights and results rather then the process of building, then NexusTrade is the platform for you!
Concluding Thoughts
While a mainstream LLM being built to access the web is cool, it’s not as useful as setting up your own custom assistant. A purpose-built assistant allows you to access the exact data you need quickly and allows you to perform complex analysis.
This article demonstrates that.
It’s not hard, nor time-consuming, and the end result is an AI that you control, at least in regards to price, privacy, and functionality.
However, if the main thing that matters to you is getting quick, accurate analysis quickly, and using those analysis results to beat the market, then a platform like NexusTrade might be your safest bet. Because, in addition to analyzing stocks, NexusTrade allows you to:
Create, test, and deploy algorithmic trading strategies
Browse a library of real-time trading rules and copy the trades of successful traders
Perform even richer analysis with custom tags, such as the ability to filter by AI stocks.
But regardless if you use Grok, build your own LLM, or use a pre-built one, one thing’s for sure is that if you’re integrating AI into your trading workflow, you’re gonna be doing a lot better than the degenerate that gambles with no strategy.
That is a fact.
Leave a Reply