Machine Learning and Momentum for Weekly Long-Only Asset Allocation

Summary

This project describes a long-only portfolio that reallocates capital among a small basket of sector or asset ETFs and cash on a fixed weekly schedule. It compares classifier-based allocations with a rule-based momentum baseline. The baseline averages recent daily returns for each asset, assigns weight in proportion to positive momentum, and holds cash when all candidate assets have negative momentum. For the machine-learning approach, the project considers asset selection, technical-indicator features, and the challenge of predicting the best-performing asset across multiple classes.

The author motivates diversification through cross-asset behavior, using five-day return correlations and conditional return frequencies to argue that gold may cushion declines in equity-linked holdings. The initial five-candidate basket reportedly produced poor results, prompting a smaller basket and further comparisons. The project also identifies limits and possible extensions, including heuristic momentum parameters, feature and model overfitting, and the exclusion of transaction costs and slippage from rebalance-frequency optimization. Its reported experiments are project-specific and do not establish performance across other periods or asset universes.

Key ideas

• The portfolio reallocates capital among a selected basket and cash at weekly intervals.
• The rule-based baseline weights assets according to average historical returns and assigns no weight to assets with negative momentum.
• The project compares this baseline with machine-learning classifiers that predict asset allocation choices.
• Correlation analysis is used to assess whether gold may diversify equity-linked holdings during market declines.
• Basket choice, model complexity, transaction costs, and slippage can materially affect results.

Tags

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