Looking at occurrences through a (daily) timeframe How much distortion in our view can we expect when we look at data through a timeframe? For example, instead of stating the exact time of some item (e.g. 2020-03-23 2h33) we group and represent all cases in a daily count (2020-03-23 100 cases, 2020-03-24 121 cases, 2020-03-25 111 cases)?
By how much would this grouping lead us astray when we try to forecast the development of something?
And yes, the question popped up in my mind watching the coronavirus pandemic, but I'm interested in scenarios like this in general.
 A: The answer is a resounding "it depends".
What forecast granularity do you need? Do you actually need predictions of the timestamps of future occurrences? Then, if you work in daily buckets, you would need some way of distributing your daily forecasts over the day. (Plus, a way of including the uncertainty in disaggregated timestamps. Also, a way of propagating the uncertainty in daily totals to the timestamps.) If you have some algorithm to do this, it may well be competitive to an approach that directly predicts future timestamps (e.g., using waiting time models).
Actually, it's hard enough to even say whether one set of timestamp predictions is better than another one. Classical point forecast accuracy measures don't really work.
However, you may actually need bucketed data. For instance, if you are planning your hospital's workforce or ICU capacity, it's not all that interesting at which precise second a new patient presents, but how many new patients arrive in a given two-hour or eight-hour slot. You could either forecast on the granularity you will need in the end, or predict timestamps and aggregate (and again, aggregate the predicted uncertainty). Again, which approach is better is not clear before you actually do the exercise.
