|
Mean-variance portfolio based on covariance matrix prediction with LSTM networks using Cholesky decomposition approach
|
Seyfollah Heydari1   , Amir Mohammadzadeh *2 , Soheyla Khishtandar2 |
1- Islamic Azad University, Qazvin Branch, Faculty of Management and Accounting;
2- Islamic Azad University, Qazvin Branch, Faculty of Management and Accounting |
|
|
Abstract: (580 Views) |
| The Markowitz mean - variance model is based on minimizing risk at different expected return levels . However , the key to success in this approach relies on accurately predicting the return vector and the dynamic covariance matrix among assets , as these relationships are constantly changing , and understanding this dynamism is crucial for achieving an optimal portfolio . The present research utilizes Long Short - Term Memory (LSTM) networks for the dynamic predicted of the covariance matrix , with its innovaton lying in addressingthe issue of the predicted covariance matrix not being positive definite by employing the Cholesky decomposition approach . In this study , the mean - variance approach was used to optimize a portfolio consisting of 10 major Tehran Stock Exchange indices the period 1390 to1404 comparing three models : LSTM networks with Cholesky decomposition , Hidden Markov , and DCC - GARCH . The results showed that the LSTM model , with a weekly average return of 0.0050 , risk of 0.038 , Sharpe ratio of 0.13 , and cumulative teturn of 2.42 , demonstrated significant superiority over the other two models in controlling risk and profitability .This superiority stems from LSTMs ability to model the complex and dynamic patterns of the Iranian capital market . |
|
| |
|
|
Full-Text [PDF 1286 kb]
(176 Downloads)
|
Type of Study: Theoretical Article |
Subject:
Financial Institutions and Services (G2)
Received: 2026/04/15 | Accepted: 2026/06/1 | Published: 2026/06/23
|
|
|
|
|
|
|
| References |
|
1.
|
|
| Send email to the article author |
|
|
