While quarterly consumption data are known to be well fitted by an integrated first-order moving average process, IMA (1,1), with a positive coefficient, it is found that monthly consumption data are well fitted by the same type of process, but with a negative coefficient. This sign reversal has three main implications. First, if the random walk hypothesis of consumption behavior is true, then the agent's decision interval must be greater than a month. In particular, this evidence rejects the possibility of continuously taken decisions of Hall's type. Second, quarterly data can be indistinguishably generated by temporal aggregation of either a random walk or an IMA (1,1) process with negative coefficient. Third, if consumption decisions are generated as an IMA (1,1) process at intervals shorter than a month, then the coefficient must be negative. The theoretical effects of temporal aggregation on IMA (1,1) processes are also investigated, and some implications for empirical inference discussed.