Moving Average Without a Loop

Every latency dashboard, every smoothed training curve and every time-series feature starts here.

Implement sliding_window_mean(x, window). Given a 1-D array x of length n and an integer window, return a 1-D float array of length n - window + 1 where entry i is the mean of x[i : i + window].

x      = [1, 2, 3, 4, 5], window = 3
result = [2.0, 3.0, 4.0]

If window is larger than n, there is no complete window, so return an empty float array.

Do it without any Python loops: no for, no while, no comprehension over positions. The straightforward vectorised answer is a prefix sum, and the whole exercise is getting its two ends right.