Regime-Based Dynamic Asset Allocation

Abstract

Data-driven investment strategies have risen in popularity over the last few decades. Markowitz introduced the idea of optimizing over a mean-risk framework, and there has been a lot of work focused on developing other strategies motivated by Markowitz's formulation. In general, however, these approaches have had mixed empirical results, due to various practical considerations such as the non-normality of asset returns and the difficulty in predicting the expected return and risk of asset prices.

Recently, prior work has shown that asset returns can be characterized by different regimes since asset prices tend to exhibit persistent trends for a period of time, followed by sudden changes. A simple example is a split into bullish and bearish regimes, where price dynamics like correlations and volatilities might be dependent on the underlying regime. Applying this framework to asset allocation can generate regime-dependent dynamic portfolio allocation strategies.

In this work, I make use of discrete jump models and \(\ell_1\) trend filtering algorithms to identify market regimes, and I use macroeconomic data to predict market regimes for subsequent periods using a gradient-boosting classifier. Given the predicted regime, I then test various regime-dependent dynamic-weight allocation strategies that incorporate information about the expected regime for the next month to adjust portfolio weights. I find that one can realize significant gains with such a dynamic-weight portfolio compared to a constant-weight portfolio like the 60/40 portfolio.