Hands-On Machine Learning for Algorithmic Trading

ICA assumptions

ICA makes the following critical assumptions:

• The sources of the signals are statistically independent
• Linear transformations are sufficient to capture the relevant information
• The independent components do not have a normal distribution
• The mixing matrix A can be inverted

ICA also requires the data to be centered and whitened; that is, to be mutually uncorrelated with unit variance. Preprocessing the data using PCA as outlined above achieves the required transformations.