A multi-step and multivariate financial time series forecasting framework based on ConvLSTM optimized by Island Algorithm

Multi-scale and multi-indicator financial time series forecasting is a valuable requirement of quantitative analysis and intelligent decision-making. Nevertheless, the current research concentrates on single-step or univariate prediction, which is unable to completely meet the real-life needs of multi-step and multivariate prediction problems in complex financial time-series situations. Convolutional Long Short-Term Memory (ConvLSTM) networks combine convolutional feature extraction with temporal dependency modeling. This combination supports multivariate time series forecasting. Nevertheless, some important hyperparameters like network structure and learning rate are often tuned manually and therefore, it has low parameter configuration efficiency and high reliance on human knowledge. To overcome these challenges, this paper suggests an intelligent financial time series forecasting system that relies on ConvLSTM optimized by Island Algorithm (IA). The Island Algorithm is employed to optimize the important hyperparameters of the ConvLSTM model, including network structure, learning rate, training epochs, and convolution kernel size. By automatically searching for suitable parameter configurations, the proposed framework reduces the dependence on manual parameter tuning and improves the adaptability of ConvLSTM for financial time series forecasting tasks. In addition, three dimensions of evaluation are considered from three perspectives: forecasting performance, reliability of models, and value of practical application, which offer systematic dimensions for evaluating financial time series forecasting frameworks. The proposed method gives a systematic approach to utilize intelligent optimization algorithms to aid deep learning models in financial time series forecasting and provides references for intelligent forecasting applications in various financial scenarios.