Portfolio Tail-Risk Protection With Non-linear Latent Factors

Abstract

Tail risk protection is a mantra in portfolio allocation. A common method in this context is the NMFRB allocation. Here, we extend it to drawdown risk measures and show that the proposed portfolios compete with machine learning-based portfolios such as Hierarchical Risk Parity (HRP) and Hierarchical Equal Risk Contribution (HERC), offering potential outperformance. The basic idea is to develop a dynamic tail-risk protection strategy using a non-linear non-negative latent factor model with an autoencoder architecture to address unstable correlations and non-linear relationships in tail risk. The probability of non-activation of latent factors is modeled via an ARMA-GARCH process. Out-of-sample tests show reduced drawdowns and statistical evidence of strategy outperformance corrected for data snooping. Despite overshooting latent tail risk, the strategy improves risk-adjusted returns and could generate substantive cumulative alpha.

Keywords: Portfolio allocation, factor model, autoencoder, non-negative matrix factorization, clustering, tail risk protection

JEL Classification: C10, C14, C21, C22, C45, C58, G10, G11