Financial forecasting using computational intelligence nowadays remains a hot topic. Recent improvements in deep neural networks allow us to predict financial market behavior better than traditional machine learning approaches. In this paper, we propose three novel deep learning-based financial forecasting frameworks, all of which considerably outperform existing approaches, yielding a much better annual financial return on DOW-30 stocks and Exchange-Traded Funds (ETFs) tested between January 1, 2007, and December 31, 2016. The first framework Convolutional Neural Networks with Technical Indicator Clustering (CNN-TIC) creates images with multiple channels corresponding to the technical indicator clusters and employs the take profit and stop loss techniques to obtain a superior annual financial return. The second model Evolutionary Optimized CNN-TIC (EO-CNN-TIC) computes the optimal values in the take profit and stop loss techniques using one of the recently created evolutionary optimization algorithms, Cuckoo Search. Finally, the third model Residual Network with Technical Analysis (ResNet-TA) applies residual blocks to the convolutional part of the neural network architecture to extract more useful features from deeper layers. Both CNN-TIC and EO-CNN-TIC are based on clustering the technical indicators by their similarity in behavior and creating separate five distinct images based on the five clusters, while ResNet-TA takes advantage of going deeper in the network with residual blocks. All three models further improve their performances by hyperparameter tuning. On DOW-30 stocks, we were able to achieve annual returns of 20.45% , 29.54% , and 36.70% for CNN-TIC, EO-CNN-TIC, and ResNet-TA, whereas for ETFs, 16.56% , 19.20% , and 32.09% annual returns were observed, respectively. We conclude with future work that can be done in order to further improve the computational and financial performances of the models.

如今，使用计算智能进行财务预测仍然是一个热门话题。深度神经网络的最新改进使我们能够比传统的机器学习方法更好地预测金融市场行为。在本文中，我们提出了三个新颖的基于深度学习的财务预测框架，所有这些框架都大大优于现有方法，在2007年1月1日之间进行测试的DOW-30股票和交易所买卖基金（ETF）产生了更好的年度财务回报，以及2016年12月31日。具有技术指标聚类的第一个框架卷积神经网络（CNN-TIC）创建具有与技术指标聚类相对应的多个通道的图像，并采用止盈和止损技术来获得卓越的年度财务回报。第二种模型进化优化的CNN-TIC（EO-CNN-TIC）使用最近创建的进化优化算法之一Cuckoo Search计算获利和止损技术中的最优值。最后，带有技术分析的第三个模型残差网络（ResNet-TA）将残差块应用于神经网络体系结构的卷积部分，以从更深的层中提取更多有用的特征。CNN-TIC和EO-CNN-TIC均基于技术指标在行为上的相似性进行聚类，并基于五个聚类创建单独的五个不同图像，而ResNet-TA则利用残差块深入网络。这三种模型都通过超参数调整进一步改善了它们的性能。在DOW-30股票上，我们可以实现20.45％，29.54％，CNN-TIC，EO-CNN-TIC和ResNet-TA分别为36.70％和36.70％，而ETF分别为16.56％，19.20％和32.09％。我们以将来可以做的工作作为结束，以进一步改善模型的计算和财务性能。