基于深度学习LSTM神经网络的全球股票指数预测研究

作为深度学习技术的经典模型之一,长短期记忆(LSTM)神经网络在挖掘序列数据长期依赖关系中极具优势。基于深度神经网络优化技术,本文构造了一个深层LSTM神经网络并将其应用于全球30个股票指数三种不同期限的预测研究,结果发现:①LSTM神经网络具有很强的泛化能力,对全部指数不同期限的预测效果均很稳定;②相比三种对照模型(SVR、MLP和ARIMA),LSTM神经网络具有优秀的预测精度,其对全部指数的平均预测精度在不同期限上均有提升;③LSTM神经网络能够有效控制误差波动,相比三种对照模型,其对全部指数的平均预测稳定度在不同期限上亦均有提高。鉴于LSTM神经网络在预测精度和稳定度两方面的优势,其未来在金融预测等方向将有广阔的应用前景。

A Study on Forecast of Global Stock Indices Based on Deep LSTM Neural Network

The Long-short Term Memory (LSTM) neural network, as one of the classic models in deep learning technology, is advantageous in mining long-term dependency of sequential data. Based on optimized technology of deep neural network, this paper constructs a deep LSTM neural network to forecast 30 stock indices in three scenarios with different horizons. The results show that i) the LSTM neural network is highly capable of generalization in financial forecast and can generate stable forecasts for 30 stock indices; ii) the LSTM neural network offers high accuracy in long and short-term forecasts in comparison with other three models (SVR, MLP and ARIMA), escalating the average accuracy of all indices for different scenarios; iii) the LSTM neural network can effectively control the error fluctuations and enhances the average stability of forecasts of all indices in the three scenarios compared to the other three models. In view of the advantages in terms of the forecast accuracy and stability, the LSTM neural network for sure will be widely used in the forecast for financial market in the coming days.