A problem of constructing a stable forecasting model using feature selection methods is considered. We propose to use a multicollinearity detection criterion which is necessary in the case of excessive number of features. Model is considered stable if small changes of the feature vector entail small changes of the target output vector. Mathematical definition of the model stability is also presented. Multicollinearity problem comes from correlation between features and causes loss of stability of the model. To study the properties of the detection criterion an additional research was undertaken which led to development of Belsley method. To prove the correctness and applicability of the approach to the problem of multicollinearity both theoretical reasoning and extra experiment are provided. The proposed criterion runs an algorithm to exclude correlated features, reduce dimensionality of the feature space and obtain robust estimations of the model parameters. The algorithm is based on step-regression method. The main idea is to add and remove the features consequently according to this criterion. The Lasso and LARS algorithms were chosen as the basic ones to compare with. The computational experiment is used to study an hourly-price forecasting curve problem with the proposed and the basic (reference) algorithms. The experiment is carried out using real time series of the German electricity tariffs.