Andrew Ng (see references on GitHub: https://github.com/PacktPublishing/Hands-On-Machine-Learning-for-Algorithmic-Trading) emphasizes the distinction between performance shortcomings due to a problem with the learning algorithm or the optimization algorithm. Complex models like neural networks assume non-linear relationships and the search process of the optimization algorithm may end up in a local rather than a global optimum.

If, for example, a model fails to correctly translate a phrase, the test compares the scores for the correct prediction and the solution discovered by the search algorithm. If the learning algorithm scores the correct solution higher, the search algorithm requires improvements. Otherwise, the learning algorithm is optimizing for the wrong objective.