Against the backdrop of the rapid development of digitalization and quantitative trading in financial markets, time-series data of financial asset prices has grown exponentially. Accurate prediction of price changes has become the core support for risk management, asset allocation and investment decision-making. Aiming at the problems that traditional statistical models struggle to capture nonlinear dependencies, existing deep learning methods ignore multi-scale temporal features and dynamic cross-asset correlations, and the mining of the coupling relationship between market sentiment and prices is insufficient, this paper proposes a financial asset price prediction method based on temporal feature learning, and constructs an integrated framework of “multi-scale temporal decomposition – cross-asset correlation modeling – sentiment-price collaborative learning – incremental feedback optimization”. Experiments were carried out based on 1.26 billion daily and hourly frequency data of CSI 300 constituent stocks, S&P 500 constituent stocks and Bitcoin. The results show that the model achieves a single-day rise and fall prediction accuracy of 78.2%, a Mean Absolute Error (MAE) of 0.85, a Root Mean Square Error (RMSE) of 1.12, a cross-asset linkage prediction recall rate of 83.5%, and a processing time of 1.8 hours for one million pieces of data. The approach can be exceptionally efficient when it comes to the accuracy of predictions, its long-term stability and generalization across markets as well as being able to offer a quantitative basis of making decisions that financial market participants may rely on.
