Hands-On Machine Learning for Algorithmic Trading

Lessons learned

Some of the lessons learned from twenty years of CNN architecture developments, and especially since 2012, include the following:

• Smaller convolutional filters perform better (possibly except at the first layer) because several small filters can substitute for a larger filter at a lower computational cost
• 1 x 1 convolutions reduce the dimensionality of feature maps so that the network can learn a larger number overall
• Skip connections are able to create multiple paths through the network, and enable the training of much higher-capacity CNNs