Financial Machine Learning Research on Bars, Labeling, and Meta-Labeling

Summary

This project update describes research notebooks for financial machine learning topics, including tick, volume, and dollar bars; CUSUM event filtering; vertical barriers; and triple-barrier labels. It outlines comparisons of bar sampling using weekly count stability, return serial correlation, and a normality test. Other notebooks explore meta-labeling with an image-classification example and as a secondary model applied to trend-following and Bollinger Band mean-reversion signals.

For the reported mean-reversion example, meta-labeling reduced annualized return from 58% to 44% while lowering maximum drawdown from 24% to 12.3%. These figures are specific to the described experiment and do not establish general performance. The update also notes that transformed sample data were shared rather than the underlying tick history, and that package implementations were still being developed. Its findings are best read as an account of research workflows and preliminary examples, not evidence that meta-labeling will improve other strategies.

Key ideas

• The notebooks compare tick, volume, and dollar bars using count stability and return statistics.
• The labeling research covers CUSUM filters, vertical barriers, and the triple-barrier method.
• Meta-labeling is explored as a secondary model for both trend-following and mean-reversion signals.
• In the reported Bollinger Band example, meta-labeling lowered return and maximum drawdown.
• The reported experiment is specific to its setup, and the package work described was ongoing.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.