Appendix C. Convolutional Neural Networks
Part III focuses on dense neural networks (DNNs) and recurrent neural networks (RNNs) as two standard types of neural networks. The charm of DNNs lies in the fact that they are good universal approximators. The examples in the book for reinforcement learning, for instance, make use of DNNs to approximate the optimal action policy. On the other hand, RNNs are specifically designed to handle sequential data, such as time series data. This is helpful when trying, for example, to predict future values of financial time series.
However, convolutional neural networks (CNNs) are another standard type of neural network that is widely used in practice. They have been particularly successful, among other domains, in computer vision. CNNs were able to set new benchmarks in a number of standard tests and challenges, such as the ImageNet Challenge; for more on this, see The Economist (2016) or Gerrish (2018). Computer vision in turn is important in such domains as autonomous vehicles or security and surveillance.
This brief appendix illustrates the application of a CNN to the prediction of financial time series data. For details on CNNs, see Chollet (2017, ch. 5) and Goodfellow et al. (2016, ch. 9).
Features and Labels Data
The following Python code first takes care of the required imports and customizations. It then imports the data set that contains end-of-day (EOD) data for a number of financial instruments. This data set is used throughout the ...