Dynamic consistency is a key behavioral property in dynamic models, enabling tractability by means of dynamic programming methods. However, it is a behavioral property that is often violated in experiments. This paper shows that dynamic consistency can be relaxed to hold over a much smaller domain of consumption programs. Nonetheless, this domain can still be sufficiently rich for practical applications. To illustrate, I provide examples of domains that are rich enough to separate risk aversion from intertemporal substitution. As an application, I introduce a new model of dynamic preferences, the Epstein- Zin-Selden-Stux preferences. These preferences are recursive only within a restricted domain. In contrast with standard recursive preferences, this weaker notion of dynamic consistency allows for indifference to the timing of resolution of uncertainty.
Supplementary notes can be added here, including code and math.