The main objective of this paper is to analyze the problems associated with the study of the variables that influence the determination of the technical capacity for the insurance business. In this sense, we tests two methodologies for calculating the total expected losses in a real portfolio of car insurance. In the first step, we modeled a conventional univariate statistical model, where we found that the LogNormal distribution was the best fit. As a second methodology, a copula model was calibrated to incorporate the stochastic association between variables frequency and severity. The results showed that both distribution have extreme behaviors, with a negative and low Kendall correlation (-0.24), but that reject the hypothesis of independence at 5 % confidence. The rotated Clayton 270 degrees copula, with the marginal distributions Exp-Poisson and Log-Normal for frequency and severity respectively, presented the best estimates of the expected loss with a difference of 12 % in relation with the empirical expected loss of the real data base. Finally, we detected that the hypothesis of independence between severity and frequency for this practical exercise leads to significant overestimates of the expected loss.