One of the most important problems in business forecasting is the problem of variables selection. It arises in many contexts and in many applications and is typically solved either using cross-validation techniques or with information criteria. According to the latter approach, a researcher selects the predictive model that has the smallest information criterion value among the models considered in a pool. This approach showed its efficiency in many contexts, but it is oblivious of the performance of models several steps ahead and on the data the model has not seen (test set). We propose a new method for selecting the best predictive model, based on direct use of Bayes’ theorem. In this method, a priori and a posteriori probabilities of suitability of each model for forecasting are calculated. A coefficient is then calculated using geometric mean of these estimates. The model for which the proposed calculated coefficient has the maximum value is considered to be the most appropriate for the data and can be used for forecasting. A comparative analysis of the main methods for selecting the best forecast model carried out by the authors on the example of one hundred different time series using autoregressive models showed that on average the proposed method for selecting a model works better than the existing ones.