Portfolio Management Based on TGCN with Attention Fusion

Portfolio management serves a dual and crucial purpose: to attain a predetermined expected return at a specific future juncture and to keep asset volatility well within an individual's risk - tolerance threshold. By diversifying investments, a portfolio can not only secure relatively stable returns across various market cycles but also mitigate losses during extreme market scenarios. In the realm of the financial market, prior to investing in any asset, a comprehensive analysis of its historical data is imperative. A variety of well - established time - series processing networks have emerged as powerful tools. These networks are highly effective in extracting key time - series features, which in turn, offer valuable insights for formulating trading strategies. Nevertheless, relying solely on time - series information often falls short. In a financial portfolio, assets are not isolated entities; they are intertwined through correlations. For instance, assets within the same industry sector or those related to the same popular investment theme typically exhibit similar trends and movements. When evaluating a single asset, it is essential to consider the information of other correlated assets. Moreover, these correlations are not static; they evolve over time. To address these challenges, this paper presents a novel portfolio weight management model. This model integrates both temporal and spatial convolutions, enabling it to dynamically capture and model the changing spatial relationships between assets over time. Experimental results have clearly demonstrated that this model outperforms the benchmark models in terms of return and risk management, highlighting its potential and effectiveness in real - world financial applications.