Financial Time Series Data Forecasting based on CEEMDAN-BiGRU-BO Model

Abstract

In the quantitative investment and risk management system, financial data prediction, as a core link, its accuracy directly determines the efficiency of asset pricing and the return level of the investment portfolio. However, under the combined influence of multiple factors such as macroeconomic fluctuations, heterogeneity of investor behavior, and unexpected events, the financial market exhibits significant nonlinearity, non-stationarity, and high noise characteristics, posing severe challenges to traditional prediction models. To address this issue, this paper develops a hybrid model integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Bidirectional Gated Recurrent Unit (BiGRU) and Bayesian Optimization (BO) to boost the forecasting accuracy and resilience of financial time series data. This model first introduces CEEMDAN in the feature engineering stage to adaptively decompose the data, extract the intrinsic mode function (IMF) and residual components, effectively alleviating the non-stationarity and complexity of the data. Subsequently, in the prediction stage, a BiGRU is adopted to enhance the model's capacity to model long-term trends and sequence structures. Finally, BO is utilized to determine the optimal parameter combination under different datasets. Based on the above methods, this paper takes the index of the Shanghai stock exchange (SSEC), the index of the stock exchange of Thailand (SET) and the Korea composite stock price index (KOSPI) as empirical data for verification. The findings demonstrate that the hybrid model achieved mean absolute percentage error (MAPE) indicators of 0.188%, 0.291% and 0.185% respectively, which were significantly lower than those of other control models, verifying its stability and adaptability in the complex financial environment.