<p>The major problem associated with portfolio optimization is the multi-objective optimization, which captivated various financial professionals to establish the optimal weight of the funds that strengthens the forecasted returns of the investments. It also amends the risk mitigation factors that are associated with portfolio optimization. This amended portfolio optimization model improves the security levels of the efficient frontiers, which offer minimum risk potentials to the forecasted returns. Hence, to develop financial investments, a highly effective portfolio optimization model is required. To establish such anticipated returns, the conventional time series method is applied to improve portfolio generation. Moreover, the conventional portfolio optimization model generally depends on the utilization of prior stocks that provide the anticipated returns as the outcome for making it viable in the actual stock trading as a long-term investment. Simultaneously, the trading market is formed by numerous smaller opinions that have effects on the price and other market psychologies of the traders. Therefore, the traditional models with small investments are not significant. Additionally, the mean prior market returns are denoted as the forecasted returns, which show the impacts of the low-pass filters in the stock market psychology, which are responsible for the miscalculations of subsequent small forecasted returns. Capitalist psychology has notable impacts on the application of prior mean returns, which directly rely on the price of the individual short-term commodities in the stock market. Therefore, the estimation of total stock returns is integrated with portfolio optimization mechanisms in the financial environment. Here, a novel portfolio optimization and returns prediction system is implemented in this research work. At first, the significant stock market-related data is gathered from the standard data source and fed to the feature extraction process. Here, an Adaptive Hybrid Networks (AHyNets) is used for performing return prediction, in which the hybrid networks include the Spatio-Temporal Attention-based Sparse Autoencoder with Residual Gated Recurrent Unit (STASA-ResGRU). To enhance the effectiveness of the designed prediction framework, the parameters of hybrid networks are optimized by Improved Red-billed Blue Magpie Optimizer (IFS-RBMO) optimization model. Subsequently, the stock's risk level is forecasted during the return prediction operation. Here, the IFS-RBMO is utilized for deriving the optimal portfolio and rebalancing by identifying the current market condition. For the selection of the best asset trading strategy, the Sharpe ratio is used, which enhances the return in relation to risk. Finally, for estimating the implemented system, the performance investigations are carried out by contrasting the designed system with other related techniques. Here, the designed approach has attained 97.01, 96.97, 97.15, 97.41 and 97.17% accuracy value in terms of 1st, 2nd, 3rd, 4th, and 5th k-fold rates. Thus, the outcome of the designed approach shows the reliable improvement of the investment strategy by reducing the volatility. Thus, it helps to optimally reduce the specific level of risk in the return prediction process without any interference, navigating the stock market and enabling investors to make better decisions.</p>