NLP and ML Quant Researcher
Builds signals and research tools from filings, news and transcripts, with timing that holds up.
The job
An NLP and ML quant researcher builds signals, screens and research tools from text such as filings, news, earnings releases and transcripts, using methods from word lists and embeddings to large language models. The hard part is timing. A filing’s date is not when it became public, a vendor score may be stamped before it existed, and a language model trained on later data may already know how the story ended. Things go wrong through look-ahead, duplicated news and trying many models.
- Supports
- Research teams that use text, alternative data and machine learning
- Stages
- 6 to learn, then proof of work
- Two routes
- Free resources, or SelfMadeQuant programmes and projects
The roadmap
The same stages either way. The free route uses university courses, lectures, books and papers that are free online. The SelfMadeQuant route maps each stage to our programmes and ends in projects on real market data.
Stage 01 · Foundations
Python and statistics
The data stack and the inference behind every claim.
- Python data stack
- Statistical inference
- Regression
- Python for Data Analysis, 3rd edition (open edition) (opens in a new tab)
Wes McKinney
Learn pandas and NumPy for cleaning, joining and reshaping text and returns data.
- 18.05 Introduction to Probability and Statistics (opens in a new tab)
Jeremy Orloff and Jennifer French Kamrin / MIT OpenCourseWare
Covers estimation, hypothesis tests and Bayesian updating at the level research work needs.
- An Introduction to Statistical Learning with Applications in Python (opens in a new tab)
James, Witten, Hastie, Tibshirani and Taylor
Regression, classification and resampling, with Python labs you can rerun on financial data.
Stage 02 · Core
Point-in-time data and validation
When information became public, and validation that respects it.
- Point-in-time data
- Leakage
- Time-series cross-validation
- Leakage and the Reproducibility Crisis in ML-based Science (opens in a new tab)
Sayash Kapoor and Arvind Narayanan
A taxonomy of leakage types to check every text-based prediction pipeline against.
- Forecasting: Principles and Practice, time series cross-validation (opens in a new tab)
Rob J Hyndman and George Athanasopoulos
Shows rolling-origin evaluation, the time-series replacement for shuffled cross-validation.
- The Deflated Sharpe Ratio (opens in a new tab)
David H. Bailey and Marcos López de Prado
Corrects a backtest Sharpe ratio for the number of trials and non-normal returns.
- Advances in Financial Machine Learning (opens in a new tab)
Marcos López de Prado
Purged and embargoed cross-validation for labels that overlap in time.
Stage 03 · Core
Text as data
Turning documents into numbers, from word counts to embeddings.
- Bag of words
- Finance dictionaries
- Embeddings
- Transformers
- Speech and Language Processing (3rd edition draft) (opens in a new tab)
Dan Jurafsky and James H. Martin / Stanford
Free draft covering bag of words, embeddings and transformers from first principles.
- Loughran-McDonald Master Dictionary with Sentiment Word Lists (opens in a new tab)
Tim Loughran and Bill McDonald / University of Notre Dame SRAF
The finance-specific sentiment word lists, with notes on how they were built.
- Text as Data (opens in a new tab)
Matthew Gentzkow, Bryan Kelly and Matt Taddy / NBER
Survey of turning documents into features and the econometric pitfalls of doing so.
- FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (opens in a new tab)
Dogu Araci
How a BERT model is fine-tuned on financial text and compared with dictionary methods.
Stage 04 · Specialist
Large language models and measurement
Using LLMs as measuring instruments, and the look-ahead they carry.
- LLM annotation
- Measurement validity
- Look-ahead in LLMs
- Assessing Look-Ahead Bias in Stock Return Predictions Generated by GPT Sentiment Analysis (opens in a new tab)
Paul Glasserman and Caden Lin
Shows how an LLM's training data can leak future outcomes into a backtest.
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models (opens in a new tab)
Alejandro Lopez-Lira and Yuehua Tang
A widely cited test of LLM headline scoring for next-day returns.
- Large Language Models: An Applied Econometric Framework (opens in a new tab)
Jens Ludwig, Sendhil Mullainathan and Ashesh Rambachan
When LLM outputs are valid inputs for prediction versus estimation, and what can go wrong.
- Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning (opens in a new tab)
Naoki Egami, Musashi Hinck, Brandon Stewart and Hanying Wei
Corrects downstream estimates when LLM annotations contain errors, using a small hand-labelled sample.
Stage 05 · Specialist
Machine and deep learning
Models that generalise on noisy financial labels.
- Regularisation
- Neural networks
- Finance-aware validation
- Stanford CS229: Machine Learning, Spring 2022 (opens in a new tab)
Stanford Online
Lectures on supervised learning, regularisation and the bias-variance trade-off.
- Deep Learning (opens in a new tab)
Ian Goodfellow, Yoshua Bengio and Aaron Courville
Reference for neural network training, regularisation and sequence models.
- Understanding Deep Learning (opens in a new tab)
Simon J. D. Prince
Recent, clearly illustrated treatment of transformers and modern training practice.
- Empirical Asset Pricing via Machine Learning (opens in a new tab)
Shihao Gu, Bryan Kelly and Dacheng Xiu
Benchmark comparison of linear models, trees and neural nets for predicting stock returns.
Stage 06 · Specialist
Event studies
Did the text move the price, and when?
- Event windows
- Abnormal returns
- Robust inference
- Causal Inference in Financial Event Studies (opens in a new tab)
Paul Goldsmith-Pinkham and Tianshu Lyu
Revisits what event-study estimates identify, using modern causal-inference tools.
- Kenneth R. French Data Library (opens in a new tab)
Kenneth R. French / Dartmouth Tuck
Daily factor returns for building market and factor-model benchmarks for abnormal returns.
- Event Studies in Economics and Finance (opens in a new tab)
A. Craig MacKinlay / Journal of Economic Literature
The standard reference for estimation windows, abnormal returns and test statistics.
Stage 07 · Proof of work
Does 10-K tone move prices?
Pull 10-K filings from EDGAR, score their tone with the Loughran-McDonald word lists and with FinBERT, and run an event study of abnormal returns around each filing date. It answers whether filing tone carries information the market had not already priced, and whether the transformer beats the dictionary.
Stage 01 · Foundations
Python and statistics
The data stack and the inference behind every claim.
- Python data stack
- Statistical inference
- Regression
- Python for Quantitative Finance
- Scientific Python and data
- Probability, Statistics and Financial Data
- Statistical inference
Stage 02 · Core
Point-in-time data and validation
When information became public, and validation that respects it.
- Point-in-time data
- Leakage
- Time-series cross-validation
- Quantitative Research, Data and Backtesting
- Hypotheses and Point-in-Time Data
- Labels and Validation Design
Stage 03 · Core
Text as data
Turning documents into numbers, from word counts to embeddings.
- Bag of words
- Finance dictionaries
- Embeddings
- Transformers
- Text as Data and NLP for Finance
- Text representation
- Finance lexicons and disclosures
- Embeddings and classification
Stage 04 · Specialist
Large language models and measurement
Using LLMs as measuring instruments, and the look-ahead they carry.
- LLM annotation
- Measurement validity
- Look-ahead in LLMs
- Text as Data and NLP for Finance
- LLM annotation and measurement validity
Stage 05 · Specialist
Machine and deep learning
Models that generalise on noisy financial labels.
- Regularisation
- Neural networks
- Finance-aware validation
- Financial Machine Learning
- Finance-aware training and validation
- Regularization and Model Selection
- Neural Networks and Deep Learning
Stage 06 · Specialist
Event studies
Did the text move the price, and when?
- Event windows
- Abnormal returns
- Robust inference
- Event Studies and Causal Measurement in Finance
- Event-study design
- Abnormal-return measurement
- Dependence and robust inference
Stage 07 · Proof of work
Build the projects
Each project is built on real market data and reports what the data said, including when the idea failed.
- FDFinancial filing change and deterioration monitorThe ratio model put deteriorating firms at the top 61% of the time against a 23% base rate. Text-change screens did worse than random.
- ASAgentic AI Screening for Portfolio InvestmentA published LLM agent stock screen rebuilt on free data earned a Sharpe of 0.18 against 0.51 for the unscreened book over 2015 to 2023.Free with an account
- ICAgent Investment CommitteeA five-seat point-in-time voting committee on 100 S&P 500 names earned a Sharpe of −0.16 against 0.02 for no screen over 2018 to 2023.Free with an account
Practise the interview
Questions this role gets asked, each with a hint and a worked solution.
- How would you test a news-embedding feature for next-day returns while controlling for duplicated stories and timing?
- A vendor’s daily sentiment score is labelled with the trading date. What evidence do you need before backtesting it for that day’s open?
- Two forecasting models predict the same returns. How do you decide whether one really improves on the other?
Start the SelfMadeQuant route free: a free account opens one complete topic from each primer and the interview question bank. No card needed.
Start free