Classifying Optimal Pairs-Trading Rebalance Frequency with Machine Learning

Summary

The study describes a machine-learning approach for classifying the optimal rebalance frequency of a two-asset portfolio strategy. It uses minute-level prices for 50 Binance crypto assets from 2022 and 2023, forms asset pairs, and groups them by correlation. Simulated pair strategies provide the target rebalance-frequency classes, while return-distribution statistics, value at risk, and pair correlation serve as model inputs. Six supervised classifiers are compared using training and test splits with repeated bootstrap sampling.

Random forests perform best for positive- and negative-correlation pairs, while logistic regression and support vector machines lead for weakly correlated pairs; naive Bayes performs worst. The reported validation on early 2024 data finds the strongest classification for negative-correlation pairs and weakest for positive pairs. The summary also reports that the strategy beat passive holding for many monthly comparisons and all annual comparisons. These findings are limited to the assets, period, simulation design, and metrics described; the supplied text gives no detailed transaction-cost analysis or full model performance figures.

Key ideas

• The study predicts rebalance-frequency ranges from pair statistics and correlation using six supervised learning methods.
• It groups asset pairs into positive, weak, and negative correlation categories.
• Random forests lead for positive and negative pairs, while logistic regression and support vector machines lead for weak pairs.
• Classification performance is reported as strongest for negative-correlation pairs and weakest for positive-correlation pairs.
• The reported strategy comparisons favor active rebalancing over passive holding, with results dependent on the study setup.

Tags

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