Prediction results of 000001.SZ by each model.

To address the limitations of existing stock price prediction models in handling real-time data streams—such as poor scalability, declining predictive performance due to dynamic changes in data distribution, and difficulties in accurately forecasting non-stationary stock prices—this paper proposes an incremental learning-based enhanced Transformer framework (IL-ETransformer) for online stock price prediction. This method leverages a multi-head self-attention mechanism to deeply explore the complex temporal dependencies between stock prices and feature factors. Additionally, a continual normalization mechanism is employed to stabilize the data stream, enhancing the model’s adaptability to dynamic changes. To ensure that the model retains prior knowledge while integrating new information, a time series elastic weight consolidation (TSEWC) algorithm is introduced to enable efficient incremental training with incoming data. Experiments conducted on five publicly available datasets demonstrate that the proposed method not only effectively captures the temporal information in the data but also fully exploits the correlations among multi-dimensional features, significantly improving stock price prediction accuracy. Notably, the method shows robust performance in coping with non-stationary and frequently changing financial market data.

针对现有股价预测模型在处理实时数据流时存在的诸多局限——包括可扩展性欠佳、因数据分布动态变化导致预测性能下滑、难以精准预测非平稳股价——本文提出一种基于增量学习的增强型Transformer框架（IL-ETransformer），用于在线股价预测。该方法借助多头自注意力机制，深度挖掘股价与各类特征因子间复杂的时序依赖关系。此外，引入持续归一化机制以稳定数据流，提升模型对动态变化的适配能力。为确保模型在整合新信息的同时保留先验知识，本文提出时序弹性权重巩固（Time Series Elastic Weight Consolidation，简称TSEWC）算法，以实现对传入数据的高效增量训练。在五个公开数据集上开展的实验结果表明，所提方法不仅能够有效捕捉数据中的时序信息，还能充分挖掘多维特征间的关联关系，显著提升股价预测精度。尤为值得注意的是，该方法在处理非平稳且频繁变化的金融市场数据时展现出优异的鲁棒性。