Each boosting iteration aims to reduce the training loss so that for a large ensemble, the training error can potentially become very small, increasing the risk of overfitting and poor performance on unseen data. Cross-validation is the best approach to find the optimal ensemble size that minimizes the generalization error because it depends on the application and the available data.

Since the ensemble size needs to be specified before training, it is useful to monitor the performance on the validation set and abort the training process when, for a given number of iterations, the validation error no longer decreases. This technique is called early stopping and frequently used for models that require a large number of iterations and are prone to overfitting, including deep neural networks.

Keep in mind that using early stopping with the same validation set for a large number of trials will also lead to overfitting, just to the particular validation set rather than the training set. It is best to avoid running a large number of experiments when developing a trading strategy as the risk of false discoveries increases significantly. In any case, keep a hold-out set to obtain an unbiased estimate of the generalization error.