A Lightweight Multi-Perspective Fusion Framework for Stock Price Forecasting

Stock price prediction is a fundamental problem in the financial domain. Stock price datasets are prone to high noise, non-stationarity, and complex nonlinearity, creating considerable problems in terms of the robustness and generalization of prediction models. Although deep learning models like RNNs and Transformers have made some progress in recent years, they are essentially still hindered by two main drawbacks: First, the model parameters are massive and the computation burden is high, which makes it difficult to deploy the model in a resource-limited environment, such as a mobile terminal or an environment of high-frequency trading; Second, the analysis perspective is limited; most methods only consider the temporal domain model, which cannot effectively use the global periodic trend and non-local inter-sample dependency to further improve the prediction performance. In this regard, this paper proposes a light-weight multi-enhancement-pathway neural network architecture (MobileNetV3- FTG) designed to realize multi-perspective collaborative temporal modeling at extremely low computational costs. The model presents three enhancement pathways: temporal domain, spectral domain, and graph structure. The temporal enhancement pathway employs an improved architecture of MobileNetV3 to efficiently extract local temporal features by means of depthwise separable convolutions and channel attention mechanisms; the frequency enhancement pathway converts the sequence to the spectral enhancement pathway by means of Fast Fourier Transform (FFT), learns the low-frequency trends, and filters out the high-frequency noise, then reconstructs into the temporal structure by means of inverse transform; the structural enhancement pathway maps the sequence to the graph structure with temporal edges and K-nearest neighbor edges, then learns the non-local dependencies between time points by means of Graph Convolutional Network (GCN). The outputs from the three enhancement paths are dynamically weighted and fused together by an adaptive fusion module that produces complementary enhancement of information. MobileNetV3-FTG is demonstrated to outperform the baseline model (GCN-CNN-LSTM) by around 64% in terms of the prediction error (MSE) and reduce the number of parameters by around 41% on the Google stock price dataset; compared to Transformer, the parameter count is reduced by 88.6%, and the accuracy is significantly higher. This model achieves a good balance between accuracy and efficiency, providing a feasible and efficient solution for financial time series prediction in resource-constrained scenarios.