Abstarct: Given the noisy, non-stationary, and regime-dependent nature of financial market data, designing methods that possess both accurate prediction capabilities and the ability to translate signals into stable trading behavior is among the fundamental challenges in financial research. Despite advancements in stock return prediction using novel models, existing studies have largely neglected the modeling and control of trading system behavior and operational stability in the face of data noise. This research, by developing an integrated frxamework, seeks to address this deficiency through the simultaneous management of data noise, understanding market regimes, and controlling operational risk in the trading decision-making process. Therefore, this dissertation, with a dual approach, focuses on developing two independent yet aligned frxameworks in the domain of prediction and trading decision-making. In the first section, the H4CSAE prediction-oriented model baxsed on stacked autoencoders is presented. In this structure, data is first compressed using a variational autoencoder for noise reduction, and then high-level features are extracted via a convolutional autoencoder. These features are fed into a hybrid model comprising XGBoost, LSTM, GRU, and BiLSTM to predict the price movement direction over a ten-day horizon. Empirical results indicate that the proposed model, with an accuracy of 77.78%, demonstrates superior performance compared to baxseline models and significantly reduces the false positive error rate. In the second section, the MRT-R trading system is designed as a non-reinforcement learning-baxsed approach. By leveraging market regime detection, understanding the overall market context, and dynamic control rules, it converts price signals into risk-aware trading decisions. The primary approach of this section is to establish a stable balance between return and risk, control drawdown, and maintain behavioral consistency under diverse market conditions. Overall, this research demonstrates that the simultaneous development of deep feature extraction mechanisms and structured risk-aware decision-making can provide an efficient frxamework for addressing the complexities of financial markets