We compare different designs that have been used to test for an impact of time horizon on discounting, using real incentives and two representative data sets. With the most commonly used type of design we replicate the typical finding of declining (hyperbolic) discounting, but with other designs find constant or increasing discounting. As a whole, the data are not consistent with any of these usual candidate discounting assumptions, and they also imply a violation of transitivity. The results have implications for interpreting previous evidence, and pose an important puzzle for understanding inter-temporal choice.