The observation window is the span of days between the two observations Upwork Insights uses to compute a run-rate: the newest read of the page and an earlier read at least 14 days older, preferably within the previous 90 days.
The window is chosen from the series of observations by a fixed rule. Starting from the newest read, the index looks for the oldest read that is both at least 14 days earlier and no more than 90 days earlier. If there is one, that is the base, and the window is up to about three months long. If there is none, because the page was last read long ago, the index falls back to the most recent read that is at least 14 days old, and the window can be a year or more. If no read is 14 days old, there is no window and the run-rate is pending.
The minimum exists because Upwork's counter moves in lumps as contracts pay out; two reads a day apart would annualise one milestone into a fortune. The 90-day preference exists so that the figure describes the recent past, not the account's whole life. Every agency page states the window beside the run-rate: the two dates and the number of days.
A long window is not wrong, only coarser. A run-rate measured over 400 days from an archived capture is an honest average of that period and says less about the last quarter than a 60-day window would. The window shortens on its own as the page is re-read.
How Upwork Insights uses it
The window length is printed with every run-rate on an agency page, and claimed pages are re-read weekly so the window stays close to the 90-day preference. Archived captures supply the base for agencies observed live only once; method in methodology, section 2.