LLMs for Quant Research: Build Text Signals & Research Agents You Can Trust
Use LLMs to read filings and earnings calls at scale, without leaking the future into your backtest.
LLMs can read every earnings call, 10-K and news story faster than any analyst. But in quant research, a fast wrong answer is worse than none. Pretrained models have already seen the future your backtest covers. Hallucinated numbers look exactly like real ones. And an agent that writes its own backtest can quietly leak tomorrow’s prices into today’s signal.
In this 4-hour hands-on workshop, Bruno Gonçalves shows you how to build LLM research tools and prove whether they add signal. You’ll turn earnings calls into point-in-time features, build a filings Q&A system that cites its sources, and give a research agent real tools, then catch the bugs it introduces.
Then comes the part most courses skip: golden datasets, incremental-information tests and walk-forward backtests that tell you whether any of it belongs near a portfolio.
What you will get:
Full HD event recording
Certificate of participation
Slide deck PPT
A working repo with four notebooks: earnings-call feature pipeline, filings RAG app, research agent and evaluation harness
A golden dataset of about 200 hand-labelled extraction examples to benchmark your own models
A pre-deployment checklist for any LLM feature: leakage, grounding, incremental value, reproducibility
Why you should attend:
Learn where LLMs actually help. Separate realistic uses (extraction, research, coding) from the ones that burn money (price prediction, autonomous trading).
See the failure cases first-hand. Look-ahead bias, hallucinated figures, mixed-up fiscal periods and silent agent bugs, each demonstrated live and then fixed.
Build, don’t just watch. Every module ends with code you run yourself on public SEC filings and transcripts.
Take the rigor home. The evaluation patterns apply to any text signal, any model and any provider.
Run it your way. Use any major LLM API, or an open-weight local model if your firm can’t send data outside.
What you will learn:
Text as alpha: extracting sentiment, guidance changes and risk-factor diffs, benchmarked against FinBERT and Loughran-McDonald
RAG over filings: chunking long documents, parsing tables, and returning answers with citations back to the source page
Research agents with guardrails: tools for market data, filings and a Python sandbox, with read-only access, cost caps and human approval
Evaluation that holds up: golden datasets, incremental-information tests against known factors, and walk-forward backtests with multiple-testing corrections
Production discipline: pinned models, cached outputs and audit trails for model-risk review
About the speaker
Bruno Gonçalves is the Founder of Data For Science: an author, trainer and consultant specialising in Generative AI and Machine Learning. He’s previously held a tenured faculty position at Aix-Marseille Université and served as a Data Science Fellow at NYU’s Center for Data Science. Bruno holds a PhD in the Physics of Complex Systems. Bruno has taught thousands of practitioners through O’Reilly, Pearson and live workshops. In his substack he explores the concepts of production level agents from first principles.