Labelling is a design decision with more impact than model choice. Predicting the sign of the next day's return, the return over the next five days, or whether a 2% target is reached before a 1% stop are three different problems with different achievable accuracies.
Fixed-horizon labels have a known weakness: they ignore the path. A five-day label of +1% looks identical whether the trade was calm or went 6% against you first, so a model trained on it cannot learn to avoid trades you would have been stopped out of. The triple-barrier-method fixes this.
Labels built from forward returns also overlap, so consecutive rows share future data. That inflates apparent accuracy and requires purged-cross-validation with an embargo-period to evaluate honestly.
Related: triple-barrier-method, purged-cross-validation, data-leakage, meta-labelling