This study investigates the effect of non-normality on the risk of a portfolio including real estate. It was shown that the real estate as well as the financial time series under study are not Gaussian. In this framework of heavy-tailed distributions, copula functions are more suitable for modelling tail dependence than correlation coefficients. In a first step, we focus on modelling the relationship between financial assets and real estate returns for the US and Swiss market using copula functions. We find that the dependence structure between a mixed-asset portfolio (bonds and equities) and securitized real estate index is modelled by the Gumbel copula for the Swiss case and by the Clayton for the US one. The inclusion of securitized real estate into a mixed-asset portfolio does not reduce the risk contrary to the direct real estate case. The Gaussian assumption made on the marginal distributions and on the dependence structures by most financial models leads to misleading inferences in terms of risk and biases the optimal level of diversification. In a second step we quantify the over- or underestimation of the risk calculated with copulas compared to a multivariate normal distribution separating the effects of the marginal distributions and of the dependence structure. These findings have important practical consequences for asset allocation decisions.