We run an experiment to elicit preferences over state-contingent timed payouts. We analyze the data using a new revealed preference method (building on Nishimura et al., 2017 ) that can test for consistency with utility functions that increase with a given preorder. We find that correlation neutrality, a property implied by discounted expected utility, is widely violated and there is, instead, strong evidence of intertemporal correlation averse behavior. Our results suggest that utility is not additive across both states and time and that credible models of choice need to allow people to prefer negative correlation in timed payouts.