Machine learning in finance
Where machine learning helps on financial data, where it overfits, and the protocols that tell the two apart.
What this shelf covers
Machine learning meets two obstacles on market data: a low signal-to-noise ratio and labels that overlap in time. These notes cover the protocols built for both, purged and embargoed cross-validation, sample weights by label uniqueness, the triple-barrier method and meta-labeling, along with probability calibration for position sizing and the evidence on when large models help. Each note states what the method controls and which leak remains open.
9 notes
A Backtesting Protocol for the Machine Learning Era
Can Machines Learn Finance? The Sober Answer
Conformal Prediction: Intervals With a Guarantee and No Distributional Assumption
Deep Learning on the Order Book, and Why It Does Not Apply Here
Machine Learning in Finance: Which Textbook Answers Which Question
The Virtue of Complexity Debate: More Parameters Than Observations
Two Cultures, and No Free Lunch
Variable Importance Is Biased, and Permutation Extrapolates
When Cross-Validation Is Valid on Time Series, and When It Is Not