Stock Price Forecasting Based on CEEMDAN-AM-BiLSTM Hybrid Model

Abstract

The fluctuation of stock prices is closely linked to a country's economic development. However, due to the significant non-linear and non-stationary characteristics of price fluctuations, traditional prediction methods struggle to capture their underlying patterns. A CEEMDAN-AM-BiLSTM prediction model is established to predict stock prices in this paper. The original data series is first decomposed to obtain Intrinsic Mode Functions (IMFs) and residual terms by using CEEMDAN. These components are then classified into high-frequency disturbance terms and low-frequency non-disturbance terms based on the spectral characteristics of IMFs. The attention mechanism is employed to identify and focus on key IMF components, which are subsequently input into a BiLSTM network for predicting non-disturbance terms. The prediction results of each IMF component are merged to derive the final predicted value. An empirical study using the minute-level closing prices of the CSI 300 index is conducted, with comparisons made against the traditional BiLSTM model and CEEMDAN-BiLSTM model. The results show that the proposed model achieves higher accuracy in high-frequency closing price prediction and is more effective in capturing the complex features of high-frequency financial data, providing a new methodological reference for improving the precision of financial market trend forecasting.