We propose a novel Asset Allocation Forest (AAF) model that combines the well established

machine learning tool with the conventional portfolio optimization method.

The determination of locally optimal portfolio weights, which dynamically respond to

market conditions, effectively captures market regimes, structural breaks and smooth

transitions in a data-driven manner. We illustrate the proposed model using a multiasset

portfolio consisting of equities, bonds, credit, high yield and commodities. The

AAF consistently outperforms established benchmarks, including the Hidden Markov

Model (HMM), even when trading costs are taken into account. It also opens the door

to valuable economic insights. By constructing accumulated local effects (ALE) plots,

we find evidence of flight-to-safety, suggesting a strategic shift from riskier assets to

less volatile bonds during periods of increased market turbulence. Furthermore, our

model shows a pronounced preference for bonds in inflationary periods, demonstrating

its adaptability to different economic conditions.