Financial data science combines massive datasets, algorithms that learn from data, and machines that act on patterns at scale. The exam tests whether you can match each technique to its proper job.
Big data describes datasets too large, fast, or varied for traditional databases. The 4 Vs framework:
- Volume. Terabytes to petabytes.
- Velocity. Streaming (tick data, transactions, social posts) versus batch.
- Variety. Structured (tables) versus unstructured (text, images, audio).
- Veracity. Quality and reliability of the source.
KEY: Structured data lives in tables. Unstructured data (news, transcripts, satellite images) requires preprocessing before a model can read it.
Alternative data is non-traditional information used to inform investment decisions: satellite images of parking lots, credit card transactions, geolocation pings, web-scraped prices, social sentiment, and environmental, social, and governance (ESG) scores. The exam favors examples that show alternative data forecasting traditional key performance indicators (KPIs), such as foot traffic predicting same-store sales.
Machine learning (ML) uses algorithms that learn patterns from data, rather than a programmer writing every decision rule by hand. Two terms come up constantly.
Common mistakes
- Calling clustering supervised. Clustering has no target variable. If no label column exists, it is unsupervised. Trap: "k-means classification."
- Confusing AI with ML. AI is the umbrella. ML is a subset. Deep learning is a subset of ML. Trap: "ML includes AI."
- Maximizing in-sample accuracy. Out-of-sample generalization is the goal. A model nailing training and bombing test has overfit. Trap: picking the model with the highest training R-squared.
Bottom line
- Big data has 4 Vs: volume, velocity, variety, veracity. Alternative data (satellite, social, transactions) is the investment-management subset and workhorse
- Supervised ML needs labeled data and predicts known targets; unsupervised ML finds hidden structure in unlabeled data (clustering, dimension reduction)
- Regression predicts continuous outcomes; classification predicts discrete categories (default vs. no-default)
- Overfitting memorizes noise and kills out-of-sample performance. Defenses: train/validation/test splits, cross-validation, regularization
Exam shortcut
If the training data has labels, the answer is supervised. If it does not, the answer is unsupervised. For overfitting questions, the fix always involves out-of-sample validation (cross-validation, holdout, regularization). For "which AI tool reads text," the answer is NLP, every time. CFA Institute does not endorse, promote, review, or warrant the accuracy or quality of the products or services offered by FreeFellow LLC.
The full lesson (about 2,467 words, 16 min read) adds 5 worked examples, all 6 common mistakes, a self-check, free in the app.
Learning objectives
- introduction to financial data science
Browse all free CFA Level I lessons or jump into free CFA Level I practice questions.