Per https://stats.stackexchange.com/a/204977/384097 the suggestion was made that, given a known distribution, applying a cutoff as means of dealing with outliers is not data leakage. I feel uncomfortable with this unless there is a perfect way to ensure autocorrelations are not significant.
As an example given [1,2,2,3,3,3,4,4,4,4,5,5,5,6,6,7] we could then generate for each value a count [1 2 3 4 3 2 1] and may choose to throw away the 1 counts yielding [22333444455566].
Suppose the following:
By looking at the values, we may suspect that there is an upward trend with oscillations of decreasing frequency as we move away from 4 in both time directions. So context may say the trend is generally more monotonic the further from 4.
Question:
Suppose we remove low-frequency values from both sides of the distribution [1 2 3 4 3 2 1], ie resulting in frequencies [23432] ie [22333444455566]
Can we say that this removal of {1,7}, if our suppositional structure is not examined, results in a problematic peeking by removal of 1, but not in removal of 7 since rarity is measured w/r/t other values, and for 1 all those other values are known to be at later times? Whereas looking at 7, even if we know of a generally monotonic structure in time, is a look backward?