Hands-On Machine Learning for Algorithmic Trading

Simplified split-finding algorithms

The gradient boosting implementation by sklearn finds the optimal split that enumerates all options for continuous features. This precise greedy algorithm is computationally very demanding because it must first sort the data by feature values before scoring the potentially very large number of split options and making a decision. This approach faces challenges when the data does not fit in memory or when training in a distributed setting on multiple machines.

An approximate split-finding algorithm reduces the number of split points by assigning feature values to a user-determined set of bins, which can also greatly reduce the memory requirements during training because only a single split needs to be stored ...