Pruning Rare Classes in Financial Machine Learning
Summary
This document asks what Lopez de Prado means by removing observations associated with rare labels in a machine learning workflow. It contrasts the idea with common financial labeling setups: binary go/no-go meta-labels and three-way short, flat, or long labels. The quoted description says a recursive procedure removes observations from classes below a minimum frequency threshold, stopping when two classes remain.
The discussion identifies a reasonable point of confusion: with only two classes, further pruning would be unnecessary, and a three-class setup offers little room for recursion. However, the document poses the question rather than resolving it. It does not identify additional label types, explain how they might be generated, or provide examples or evidence about the procedure’s performance. Readers should treat it as a question about interpreting a book excerpt, not as a complete account of rare-class handling or a recommendation to discard data.
Key ideas
- The document describes recursively removing observations from classes whose frequency falls below a chosen threshold.
- It compares this procedure with binary meta-labels and three-way directional labels.
- The author questions why a recursive procedure is needed when only a few labels are available.
- The document does not answer what additional label types the book may have in mind.
Tags
Full text
# In Lopez de Prado's Advances in Financial Machine Learning, what is meant by "unnecessary labels"? # In Lopez de Prado's Advances in Financial Machine Learning, what is meant by "unnecessary labels"? In Lopez de Prado's Advances in Financial Machine Learning, Chapter 3, Prof. Lopez de Padro talks about dropping rare labels: ``` Some ML classifiers do not perform well when classes are too imbalanced. In those circumstances, it is preferable to drop extremely rare labels and focus on the more common outcomes. Snippet 3.8 presents a procedure that recursively drops observations associated with extremely rare labels. Function dropLabels recursively eliminates those observations associated with classes that appear less than a fraction minPct of cases, unless there are only two classes left. ``` I fail to see what could that mean, seeing as (if I understood correctly), the labels are either Go/No Go (for metalabeling) or Short/Flat/Long for "prime" labeling. Dropping rare labels until only two are left is meaningless for Go/No Go, and as for Short/Flat/Long labeling - while it is possible to apply the method and drop one of the classes if too rare, it seems to me the the implication of the language of the quote suggests that the author expects more then these three labels (otherwise the language of the quote would probably refer to it) - to speak nothing of the fact that an expectation of three labels, only one of which could be dropped would render the use of a recursive approach irrelevant). So I can summarize that the author expects different types of labels, of which more then 3 are expected. What are those labels and how are they generated?
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.