What is an appropriate method to compare prediction accuracy between different granularities in time-series data? I am doing time-series multi-variate regression on a usage variable for many individuals in a population (one model for each individual).  In some cases the individual's usage is regular enough to make predictions for each hour, but sometimes their hourly usage is so erratic that it no longer makes sense to make hourly predictions and I would rather use a model that makes predictions of the individual's daily usage.  
My current implementation is to use a threshold on the prediction accuracy for the hourly models to decide if they are good enough, and if not switch to the daily model.  I am curious though if there is an appropriate metric or technique to compare the hourly vs daily model for each individual directly.  For example, does it make sense to calculate a normalized error metric for the two models and compare them directly?
 A: This is an interesting question.
My approach would be to come to this from a different angle. You are making predictions. You probably have some reason for this: your predictions will likely drive some decisions, like provisioning some service for the required usage.
But then your subsequent decision process requires forecasts on some specific granularity. If you are forecasting for supermarket demands, you don't care about hourly granularity, only about daily (except for the deli counter, where you do need hourly forecasts). If you forecast for call center staffing, you will likely need hourly forecasts, or even forecasts on 15-minute intervals. (Incidentally, such decisions also require forecasts for the aggregate over all users, not every user separately, like your application seems to do - there are multiple aggregation dimensions to think about.)
And if your decision process dictates the time granularity, then that's really the end of the discussion, in a sense. If you need hourly forecasts, then it won't do you any good to point to good daily forecasts. Conversely, if you need daily forecasts, then good hourly forecasts are not that useful, except for being good raw material for aggregation to days.
Overall, you have a hierarchical problem. You need forecasts on some level of the hierarchy (hourly or daily), and you can calculate them in different ways: directly (forecast the hourly or daily data), bottom-up (forecast hourly, then aggregate forecasts), top-down (forecast daily, then disaggregate in some way), or even a favorite of mine, optimal combination approaches. And of course you can compare the accuracy of direct vs. bottom-up vs. optimal daily forecasts, or of direct vs. top-down vs. optimal hourly forecasts - depending on the granularity you need. In this case, you won't be comparing apples and oranges any more.
Then again, you write that some users' usage is very erratic. If you do need hourly forecasts on a single user level, then hierarchical approaches will likely not help very much in such a situation. You may want to take a step back and think about what you really need here. Do you really require each separate user's usage to be forecasted, or can you aggregate them usefully? Are you really interested in mean forecasts, or are you more interested in getting high quantiles right, to ensure a certain service level? Do you actually need forecasts for your erratic users at all, or does it make sense to just reserve enough capacity to serve them when they do arrive - or conversely, intervene to make their usage either more forecastable (like by requiring advance notice, essentially moving from a make-to-stock scenario to a make-to-order one), or to make them pay a premium for the supply chain problems their erratic behavior causes?
