Temporal Multivariate Mixture Density Networks for Portfolio Optimization
39 Pages Posted: 22 Nov 2024 Last revised: 16 Dec 2024
Date Written: October 17, 2024
Abstract
This study capitalizes on recent advancements in deep learning and statistical analysis to propose
a novel methodology for forecasting asset price returns and quantifying associated uncertainties
using a Long Short-Term Memory (LSTM) network-powered constrained Mixture Density
Network (MDN). Specifically, three models are introduced: the LSTM Gaussian-t Mixture
Network (LSTM-GTN), LSTM Gaussian Mixture Network (LSTM-GMN), and LSTM Student-
t Network (LSTM-TN). Notably, LSTM-GTN and LSTM-GMN are designed to differentiate
between periods of market volatility and stability, offering interpretable parameters to inform
prudent investment decisions.
The preliminary model selection (PMS) process is guided by three evaluation criteria: the
lowest negative log-likelihood, the highest Sharpe ratio, and the best F1-score. The study further
develops four single-stock investment strategies based on signals derived from the parameters
of the selected models. Following this, investment model evaluation (IME) is performed by
assessing the strategies across six assets, using the average rolling Sharpe ratio to validate
the models. The most robust models are subsequently evaluated during a testing period, with
performance measured by cumulative returns, drawdown, and rolling Sharpe ratio. The findings
indicate that the proposed PMS and IME framework consistently identifies investment models
that outperform the "buy-and-hold" strategy in most scenarios.
Keywords: Temporal Multivariate Density Network, Attention LSTM, Student - t, Financial Time Series, Asset Management, Portfolio Management, Markowitz
Suggested Citation: Suggested Citation
Lam, Fong and Chan, Jennifer and Choy, S. T. Boris, Temporal Multivariate Mixture Density Networks for Portfolio Optimization (October 17, 2024). Available at SSRN: https://ssrn.com/abstract=4992772 or http://dx.doi.org/10.2139/ssrn.4992772
Do you have a job opening that you would like to promote on SSRN?
Feedback
Feedback to SSRN