为进一步研究深度学习在金融时间序列预测方面的应用效果并寻找预测效果更好的优化模型，本文基于 RNN、LSTM、Transformer 模型，并将加入注意力机制后的 LSTM 作为 Transformer 的解码器从而提出 LSTM-Transformer 模型，结合 PCA 技术、EMD 技术利用前 20 天的股票数据预测后 1 天的股票收盘价。选择的实验数据是 1633 日的沪深300 的基本行情数据和 18 个技术指标。根据各个模型的预测结果，比较模型相对误差、趋势预测准确率，得出以下结论：
1. 进行股票收盘价的单特征预测时，LSTM-Transformer 模型的预测效果好于 LSTM、RNN、Transformer 模型，但 Transformer 预测趋势的准确率最高。将基本行情数据（收盘价、开盘价、最高价、最低价、涨跌幅和成交量）输入 RNN、LSTM、Transformer 以及 LSTM-Transformer模型可以改善模型的预测能力，其中 LSTM-Transformer 模型表现最好，Transformer 模型改善最显著。
2. 结合 PCA 技术的各个模型的预测效果均好于直接向模型输入18 个技术指标及基本行情数据，其中 Transformer 的表现最好。但整体来看，加入 18 个技术指标和 PCA 技术后，各个模型的预测效果均不如直接使用 RNN、LSTM、Transformer 以及 LSTM-Transformer 模型进行单特征预测和关于基本行情数据的多特征预测。
3. 结合 EMD 技术的各个模型的单特征预测效果较基础模型均有大幅提高，其中 EMD-LSTM-Transformer 模型的预测效果最好。

In order to further study the application effect of deep learning in financial time series data prediction and find the optimization model, this paper is based on RNN, LSTM, Transformer and uses LSTM as the decoder of Transformer and proposes LSTM-Transformer model. Combined with PCA technique and EMD technique, these models is uesed to predict the close price of the stock in the next 1 day by using the stock data in the first 20 days. The experimental data chosen are from CSI 300 for 1633 days. Comparing mse and trend prediction accuracy of each model, the following conclusions are drawn:
1.When predicting the stock close price using a single feature, the LSTM-Transformer model outperforms the LSTM, RNN and Transformer models in mse, although the Transformer model has the highest accuracy in trend prediction. Inputting basic market data (Close, Open, High, Low, Change and Volume) into the RNN, LSTM, Transformer and LSTM-Transformer models can improve the predictive capability of the models, among which the LSTM-Transformer model performs the best.
2.The predictive performance of models combined with PCA technology is superior to those directly inputting 18 technical indicators and basic market data, with the Transformer model performing the best. However, including 18 technical indicators and using PCA does not improve the predictive performance, which compares to single-feature and multi-feature only based on basic market data.
3.Compared with the basic model, it is significantly improved that the performance of the model combined with EMD in single-feature prediction, with the EMD-LSTM-Transformer model outperforming the EMD-LSTM, EMD-RNN, EMD-Transformer
models.