Hands-On Machine Learning for Algorithmic Trading

Manifold learning

Linear dimensionality reduction projects the original data onto a lower-dimensional hyperplane that aligns with informative directions in the data. The focus on linear transformations simplifies the computation and echoes common financial metrics, such as PCA's goal to capture the maximum variance.

However, linear approaches will naturally ignore signal reflected in non-linear relationships in the data. Such relationships are very important in alternative datasets containing, for example, image or text data. Detecting such relationships during exploratory analysis can provide important clues about the data's potential signal content.

In contrast, the manifold hypothesis emphasizes that high-dimensional data often lies on ...