Artificial Intelligence in Finance

Appendix B. Neural Network Classes

Building on the foundations from Appendix A, this appendix provides simple, class-based implementations of neural networks that mimic the APIs of packages such as scikit-learn. The implementation is based on pure, simple Python code and is for illustration and instruction. The classes presented in this appendix cannot replace robust, efficient, and scalable implementations found in the standard Python packages, such as scikit-learn or TensorFlow in combination with Keras.

The appendix comprises the following sections:

• “Activation Functions” introduces a Python function with different activation functions.

• “Simple Neural Networks” presents a Python class for simple neural networks.

• “Shallow Neural Networks” presents a Python class for shallow neural networks.

• “Predicting Market Direction” applies the class for shallow neural networks to financial data.

The implementations and examples in this appendix are simple and straightforward. The Python classes are not well suited to attack larger estimation or classification problems. The idea is rather to show easy-to-understand Python implementations from scratch.

Activation Functions

Appendix A uses two activation functions implicitly or explicitly: linear function and sigmoid function. The Python function activation adds the relu (rectified linear unit) and softplus functions to the set of options. For all these activation functions, the first derivative is also defined:

In [1]: import