<p>Stock data are inherently time-series data, characterized by complex structures and high levels of randomness. This makes forecasting stock prices a challenging problem in machine learning. The non-stationary nature of stock indices in financial markets, coupled with randomness and uncertainty in the data, and the simple hierarchical structure and lack of semantic information in the temporal dimension pose significant challenges to traditional time-series forecasting models. This study therefore proposes a novel stock price forecasting framework based on a hierarchical LSTM model to address these issues, which is known as Adv-SHNets. The model’s core concept involves learning the temporal features of stock data through two layers of LSTMs to fully capture its hierarchical structure. The framework processes stock time series via overlapping segmentation in the lower LSTM layer to preserve temporal continuity and enrich single-step semantics. Subsequently, a segmented attention mechanism and upper LSTM capture multi-scale hierarchical dependencies. A simple yet effective adversarial loss function significantly enhances the model’s robustness against data uncertainty. Furthermore, experiments on two real-world stock datasets demonstrate that Adv-SHNets achieves a relative improvement of 12.97% over comparable state-of-the-art methods, providing thorough validation of the effectiveness of this approach.</p>