Hands-On Machine Learning for Algorithmic Trading

How DL extracts hierarchical features from data

The core idea behind DL is that a composition of factors, or features, potentially organized in a hierarchy of multiple levels, has generated the data. Hence, a deep model encodes the prior belief that the target function is composed of simpler functions. This assumption allows an exponential gain in the number of regions that can be distinguished for a given number of training samples.

In other words, DL is a representation learning method that extracts a hierarchy of concepts from the data. It learns this hierarchical representation using neural network architectures that compose simple but non-linear functions and successively transform the representation from one level (starting with the ...