A process for modeling numerical data for forecasting a phenomenon relates to constructing a model by processing and learning on collected data. The fit and robustness of the model are evaluated and the model parameters are adjusted to select an optimal model in the form of a D^th order polynomial. A trade-off between learning accuracy and learning stability is controlled by adding to a covariance matrix a perturbation in the form of the product of a scalar times a matrix H or in the form of a matrix H dependent on a vector of k parameters =(₁, ₂, . . . _k). A data partition step can divide the data into a first subset for constructing the model and a second subset for adjusting the value of the model parameters according to a validity criterion obtained from data that was not used to construct the model.