In this paper, we empirically analyze the possible advantages of modelling a time-varying risk aversion that best fits investors’ behavior in the context of the optimal portfolio choice. We build optimal dynamic portfolios by focusing on the estimation of a time-varying relative risk aversion parameter (RRA). Conditional univariate and multivariate models, such as GARCH, GARCH-M and DCC-GARCH, for modelling the optimal portfolio choice and the RRA parameter are implemented. As a model validation tool, the realized performance and downside risk exposure of these portfolios one month ahead is compared to that resulting from implementing a constant risk aversion parameter. The Ledoit and Wolf (2008) test provides robustness to our results and reveals the average outperformance of the dynamic risk aversion strategy over others as the constant risk aversion or the passive management strategies.