A question that tends to arise in multiple regression models (with p variables), and that has not been totally satisfactory solved, is to determine the best subset or subsets of q q≤p) predictors which will establish the model or models with the best discrimination capacity. This problem is particularly important where p is high and/or where there are mutually redundant predictors. With this work, we present a new approach to this problem, where we will try to predict a new emission episode of S O2, but focusing our attention in the importance to know the best combinations of time instants to obtain the best prediction. The proposed method is a new forward stepwise‐based selection procedure that selects a model containing a subset of variables (or time instants) according to an optimal criteria (determination coefficient or Akaike Information Criterion) and taking into account the computational cost. Additionally, bootstrap resampling techniques are used to implement tests capable of detecting whether significant effects of the unselected variables are present in the model.