I have some 1-minute bar data. The first datapoint has time t0 and the last one t1. 99.5% of the data between the first timestamp and the last timestamp is there and 0.5% is missing (NaN values). I can confirm that there is a data point for every single minute between t0 and t1.
As a data pre-processing step, for consistency, rather than dropping the NaN values, I was wondering if I can substitute them with a rolling mean or median. Because there are very few missing values (0.5%) this shouldn't cause any ill effects. Would you do this? Would you use the median or mean or just forward/front fill or back fill?