High-dimensional Cross-market Dependence Modeling and Portfolio Forecasting by Copula Variational LSTM

In the increasingly connected world, many systems are more or less coupled with each other in various ways. A typical example is the cross-market portfolio management, where the products of heterogeneous markets are selected and configured for investment. In such cross-market problems, one market is coupled with and influenced by others, and the financial variables of a market are coupled over time. This work makes the first attempt to model both the observations-based and latent dependence degrees and structures of highdimensional financial variables in cross-market portfolios by integrating variational recurrent neural networks. It integrates the distribution-based sequential modeling of multivariate time series and the regular vine copula-based dependence structures for modeling variable dependencies. Our method addresses the needs and gaps of modeling non-normal and long-range distributional interactions across market variables. We verify the model in terms of both technical significance and portfolio investment performance against benchmarks including linear models, stochastic volatility models, deep neural networks, and variational recurrent networks for portfolio forecasting.