Using Cycle States and Machine Learning to Rank Asset Returns

Summary

The report describes an asset allocation approach that measures cycle states and uses machine learning to estimate the probability that assets will outperform one another. It reviews macroeconomic timing frameworks, then describes cycle factors derived from recurring market patterns. Signal processing methods are used to identify common cycles; asset year-over-year series and their fitted values and changes are used to characterize an asset’s state. Machine learning maps those features to probabilistic return rankings, while an ensemble across parameter choices is intended to reduce sensitivity to any single model configuration.

The summary reports historical allocation tests across global and Chinese markets, with the global stock and bond application covering January 2004 through March 2018. It states that results exceeded an equal-weight benchmark and gives historical return and risk statistics, but the underlying report and full methodological details are not included here. The authors caution that cycle lengths are estimates, policy shocks and short-term volatility can disrupt them, and historical relationships may fail. The reported results are sample dependent and do not establish future performance.

Key ideas

• The approach represents asset cycle states with fitted values and changes from cycle factors applied to year-over-year series.
• Machine learning estimates probabilistic rankings of future asset returns from cycle-state features.
• An ensemble across parameter settings is used to reduce dependence on a particular model choice.
• The summary reports historical tests in global and Chinese markets against an equal-weight benchmark.
• Cycle lengths are uncertain, and policy shocks or changing market relationships can weaken the method.

Tags

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