Predicting Multi-Indicator Stock Time Series using Convolutional Neural Networks based on Feature Engineering

Stock data is a typical class of time series data, which is characterised by complex formation mechanism and rich time granularity, and due to its financial attributes, it has always been a hotspot and a difficult point in time series research. Based on feature engineering. This study proposes a novel approach to predicting stock price movements using convolutional neural networks (CNNs) and feature engineering. We select key technical indicators through random forest modeling and transform multi-indicator time series into composite images using an enhanced Gram angular field method. These images are then used to train CNNs for predicting stock trends. Experiments on different time granularities of the GEM index (399006.SZ) demonstrate prediction accuracies exceeding 85% on both training and validation sets. This approach effectively captures temporal features and multi-dimensional information in financial time series, offering improved predictive performance over traditional methods. However, the model’s performance is contingent on sufficient data availability, suggesting areas for future research.