Stock portfolio optimization is an important area in financial decision-making. It focuses on optimizing the balance between risk and return. Our approach is centered exclusively on stock data, based on the assumption that market dynamics and price movements capture all relevant available information. We take advantage of the non-stationary and heterogeneous nature of financial time series by decomposing features into multiple intrinsic modes, which provide a richer representation of the temporal information and integrate with the hyper-graphs and attention mechanisms which represent complex relationships within the stocks in the portfolio. This paper presents a novel hybrid model that integrates Intrinsic Mode Functions (IMFs) derived from Variational Mode Decomposition (VMD) with Long Short-Term Memory (LSTM) networks for temporal data modeling, and a Hypergraph Attention Module (HGAM) to represent group-level dependencies among stocks. Our decision-making layer uses a simple Reinforcement Learning (RL) approach. The model is trained and evaluated using data set with IMFs replacing selected features. Experimental results indicate improved portfolio performance, evaluated by Accumulated Portfolio Value (APV), Sharpe Ratio (SR), and Maximum Drawdown (MDD), particularly when IMFs are included in the dataset replacing original selected features.