Bayesian inference for risk minimization via exponentially tilted empirical likelihood