Researcher course in machine learning
Alternative Data and NLP for Finance
Alternative data and natural language processing as funds use them: turning news, filings and earnings calls into signals with language models, judging satellite, card and web data, and testing whether any of it adds return once costs and timing are honest.
Alternative data is information beyond prices and company accounts that investors use to see changes earlier: the tone of an earnings call, a shift in a filing’s wording, spending on cards, cars in car parks. This course teaches how to turn such data into signals, especially text with natural language processing, and how to test whether a dataset is worth its price.
Why is timing the hardest part?
Because a signal is only real if the data was available when the trade would have been made. News carries publication times, filings are revised and vendors backfill history. The course builds every signal on point-in-time data so that the backtest matches what could have been traded.
How is the course taught?
Audio lessons give the general subject, explainer films show text turned into numbers and signals, and the workbooks hold notebooks, worked examples and cheat sheets. Assignments and the project receive written feedback from the faculty.
Who is this course for?
Quants, data scientists and analysts at funds that buy or build alternative data. It counts toward the Certificate in Machine Learning for Trading and so toward the Fellowship in Computational Finance.
What you will be able to do
- Turn news, filings and earnings call transcripts into features with classical NLP and language models.
- Build and test a sentiment signal with correct timestamps and no look-ahead.
- Evaluate an alternative dataset for coverage, history, bias and decay.
- Write the due-diligence note a fund needs before buying a dataset.
Syllabus, module by module
What alternative data is
Text, transactions, satellite, web and app data, and the market for them.
Text as data
Tokens, dictionaries, the Loughran-McDonald word lists and embeddings.
Language models
Transformer models and large language models for classification, extraction and summaries.
News, filings and calls
Sentiment, surprise and change in tone; timestamps and point-in-time text.
Other alternative data
Card spending, satellite counts and web traffic: coverage, bias and decay.
Due diligence
Backtesting a dataset, legal and licensing checks, and the note to the investment committee.
How the certificate is earned
Graded assignments and a final project: a text-based signal built from filings or news, tested out of sample, with a dataset due-diligence note. On a pass you receive the Course certificate in Alternative Data and NLP for Finance, issued by the Institute with a verification code. Read how certification works.
Before you start
Machine learning for trading at the Institute's level.
This course counts toward
- Certificate in Machine Learning for TradingOne semester of eight courses and a capstone on applying machine learning to markets without fooling yourself: labelling, features, purged validation, deep and language models, regimes and reinforcement learning.
- Fellowship in Computational FinanceThe top credential on the researcher path. Four semesters, 38 courses, and a dissertation supervised in writing and examined by two faculty members.
Questions applicants ask
What is alternative data in finance?
It is any data beyond prices, volumes and company accounts that investors use to forecast returns, such as news text, card spending, satellite images or web traffic.
Can language models predict stock returns?
They are good at reading text quickly and consistently, which can produce useful signals from news and filings. Much of the easy value is already priced, so the course tests every signal out of sample and after costs.