Presentation
Welcome to this full demonstration of QuantCoder FS, an AI-powered research assistant developed for quantitative strategy discovery. In this walk through, we present four integrated workflows — article retrieval, summarization, fundamental analysis, and code generation — all built on a modular architecture leveraging GPT-4o, CrewAI, and a custom PDF/NLP pipeline.
The system identifies relevant research articles based on keyword input, extracts structured trading logic from selected papers, generates executable QuantConnect code, and performs fundamental analysis — all while maintaining an inference cost of under $0.20.
This demonstration is intended as a technical showcase rather than financial advice. It illustrates the potential of AI to support end-to-end financial research workflows, enabling individual researchers and traders to construct their own fully integrated analysis pipelines.
Full transcript
Hello, this is Emilia Benett, the 11 Labs voice guiding you through a presentation of the QuantCoder FS application.
Before we dive into the demonstration, let’s first check the current OpenAI LLM usage balance to estimate the overall cost of this run. We're starting with $2.09.
We’ll begin by launching the article search workflow using the keyword “pair trading.” This is a particularly challenging query, as both terms — "pair" and "trading" — are fairly generic on their own. The sorting and filtering of articles may take a moment, but it typically completes in under a minute.
While we wait, a quick note: this project is developed and hosted on Substack, where we share ongoing updates, research highlights, and strategic insights. A special thanks to our first paid subscriber — your support drives this initiative forward. With the current setup, the platform is moving toward automated discovery of high-potential equities and trading strategies, powered by EODHD, a leading provider of financial data. EODHD delivers a broad suite of APIs covering global markets — including stocks, ETFs, bonds, Forex, financial news, and more. Now, back to the article search:
And here we are — the article list is ready. Each one is now available for download and further analysis.
Let’s now move to the summarization workflow. This particular article is a great candidate, as the trading logic is clearly articulated — making it an ideal test of our AI system’s ability to extract structured strategies from natural language. The summarization flow is domain-agnostic, built on robust NLP techniques that generalize well across disciplines. It’s powered by GPT-4o and a proprietary PDF reader tool, which together feed structured content into a CrewAI agent specialized in contextual interpretation — all without relying on finance-specific heuristics.
Quantitative finance articles typically span 10 to 20 pages, often dense with formulas and jargon. Despite that, the workflow is adaptable and can also process other document types, including earnings transcripts, industry whitepapers, and research briefs.
Now, reviewing the summary generated from this article, we can confirm that the core trading logic has been successfully extracted. Let’s go ahead and copy this summary into the code generation workflow.
While the summary could benefit from a bit of manual cleanup before being passed to the pair-coding assistant, we’ll proceed as-is for the sake of this demonstration. In a full research pipeline, this cleaning and validation step would be essential. The code generation workflow also includes a timer, so we’ll pause briefly and return once the process is complete.
And there it is — the code was successfully generated in about 40 seconds.
Next, let’s test whether the code compiles directly within the QuantConnect environment.
It’s important to note that generating a single article isn’t sufficient to validate the AI workflow. Instead, what begins here is a full testing campaign aimed at evaluating performance across a broader range of cases.
Let’s now launch the backtest and see how the generated code performs.
While we wait for the results, let’s switch over to our "Chat with Fundamentals" workflow and run a sample query — for example: "Is Tesla a good buy?"
As always, this is not financial advice — it’s a demonstration of an AI-driven research assistant. Users are encouraged to perform their own due diligence before making any investment decisions.
This workflow also uses a timer, so we’ll pause briefly and return when the analysis is complete.
Back to the backtest — the code compiled successfully, but the performance is modest: with a probabilistic Sharpe ratio close to zero and a CAGR of around 5%. Nonetheless, this represents a solid boilerplate for further refinement and optimization, particularly when examined in the context of the full article.
The fundamental analysis was also completed in under a minute.
One final note — the total cost for this full demonstration came to $2.25, which breaks down to just 16 cents for the 4 workflows.
With a fully self-contained, modular architecture and a robust data pipeline now in place, development will continue with a focus on extending data sources and improving the accuracy of alpha extraction.
Thank you for watching — and stay tuned for what’s next.
Disclaimer: Nothing presented in this video constitutes a recommendation to buy or sell any financial instrument. Trading and investing inherently involve risk. Viewers and users are encouraged to conduct their own due diligence.
S.M.L.
