Sequential Forecasting Strategies for Crypto Portfolio Allocation: A Dynamic Latent Factor Analysis Approach

Abstract

In this article, a highly flexible and fast framework for estimating and exploring the dynamic latent dependencies among main cryptocurrency log‐returns is introduced. Our approach is inspired by latent factor analysis models capturing simultaneously the persistence of volatility clustering and the presence of leverage effect in the crypto market. A recursive quasi-maximum likelihood strategy is proposed to infer the latent cross-sectional correlation structure of the data and then to estimate the parameters of the model. Our algorithm consists of recursively alternating between Kalman filtering and smoothing recursions and the expectation maximization (EM) algorithm. This produces economically interpretable estimates for the common latent factors, which can be used to predict the mean return vector and the return covariance matrix, particularly useful for stock selection and portfolio allocation problems. To illustrate the performance of our sequential strategy, the model is applied to daily log-returns of the last 5 years of the Bitcoin, Ethereum, Litecoin, Monero, Binance Coin and Waves currencies. The out-of-sample forecast encompassing tests show that the factorial approach yields better forecasts than those given by typical benchmarks. In terms of cumulative financial returns, our framework seems to provide the best performing mean-variance portfolio.