Optimization of Deep Learning Models for Non-stationary Time Series Data

This paper studies the optimization and application of deep learning models for non-stationary time series data, especially in the field of financial time series prediction. In view of the limitations of traditional models in processing non-stationary data, this paper designs a prediction framework based on the Transformer model. Through the series stabilization and destabilization attention modules, the model’s prediction ability for non-stationary time series data is effectively improved. Experimental results show that the designed model can achieve a price point prediction accuracy of up to 94.6%, and is superior to traditional methods in terms of computational efficiency, generalization ability and data processing capabilities. This study not only improves the accuracy and efficiency of financial time series prediction, but also provides new ideas and methods for deep learning modeling of non-stationary time series data.