Hands-On Machine Learning for Algorithmic Trading

Benchmarks

The following diagram on the left plots the top-1 accuracy against the computational cost of a variety of network architectures. It suggests a positive relationship between the number of parameters and performance, but also shows that the marginal benefit of more parameter declines, and that architectural design and innovation also matter.

The diagram on the right plots the top-1 accuracy per parameter for all networks. Several new architectures target use cases on less powerful devices such as mobile phones. While they do not achieve state-of-the-art performance, they have found much more efficient implementations. See references on GitHub for more details on these architectures, and the analysis behind these charts: