Since a risk premium is determined by the riskiness of an asset and the attitude of investors to risk, one of them or both may predict the future return variations. In this paper, I adopt the risk aversion as a proxy for the attitude of investors to risk, and examine its return predictability. More specifically, I propose the new method to estimate the time-series of risk aversion, and test whether this factor captures the time-varying risk premium (future excess returns) on the S&P 500 index. According to the results, time-varying risk aversion has a predictive power for future index returns even in short horizons, i.e., two, three and four week horizons, and does not lose the significance in the presence of conventional forecasting factors, including dividend yield, short rate, and variance premium. For robustness, I conduct an additional test on its Sharpe ratio prediction. The results regarding Sharpe ratios also support the relevance of the newly proposed risk aversion factor for capturing the variation of risk premiums, in particular, the changes in attitude of investors to risk.