Dergi makalesi Açık Erişim

Dynamic Optimization of Investment Portfolio Based on Self-Attention Mechanism - LSTM Model Prediction

2023    Wu, Songwei

In the era of big data and advanced computational capabilities, financial market participants are continuously searching for innovative strategies to gain a competitive edge. A potential pathway emerges in the domain of deep learning, especially with regard to the LSTM neural architectures, which are renowned for their ability to handle and make predictions based on time series data. This study delves into the utilization of LSTM for predicting stock prices, emphasizing the advantages of dynamic investment portfolios in the rapidly fluctuating market conditions. Utilizing a dynamic window approach for time series data preprocessing, a self-attention mechanism - LSTM model was designed to anticipate the tendency of annual closing prices for five stocks from 2022 to 2023, Utilizing the initial 80% of stock price data as training set and allocating the residual 20% for validation. The performance of the dynamic optimization portfolio model was assessed by dynamically adjusting the weights of the stocks based on the last 20% of the data, and was subsequently compared to actual market cumulative returns. The findings indicate not only that the LSTM model offers a commendable level of accuracy in predicting stock prices, but also that the recursive algorithm for the dynamic optimization portfolio, constrained by maximum returns and minimal standard deviation, consistently outperforms the general market.