For a moving window of thirty records, what decides whether the first twenty-nine positions produce a value at all?

## What the threshold counts A moving window of fixed length declares how far back it may reach. It does **not** declare how many records must actually be there. Those are two parameters, and the second one — the **minimum observation count** — is how many real observations a window must cover before it produces a value instead of an absent one. Early in a sequence there simply are not thirty records behind the current position. The window covers what exists: one record at the first position, two at the second, and so on. The threshold decides what happens in that region. - Set the threshold equal to the window's length, and the first twenty-nine positions hold no value; the output begins at position thirty. - Set it to one, and every position from the first holds a value — computed from one record, then two, then three. - Set it somewhere between, say ten, and the output begins at position ten, computed from ten observations, then eleven, until the window is full. ## It is not only the leading edge Two other situations reach the same threshold, and both are easy to forget because they happen where nobody is looking. 1. **Absent values inside the window.** A window covering thirty records, some of which hold no value, is covering fewer than thirty *observations*. Whether that position answers depends on the same threshold, in the middle of the data rather than at the head. 2. **A window whose length is a span of time.** On an irregular feed the number of records inside a fixed span changes at every position. A quiet stretch can push the covered count below the threshold, so the output has holes wherever the feed went quiet — which is often exactly where someone will later look for an explanation. ## The defaults differ, so read them Designs in this space do not agree here. Some produce a value only once the window is full. Some produce a value from the very first position. Some expose the threshold as a parameter of its own and some fold it into the operation with no way to change it. A pipeline ported from one to another can silently gain or lose a block of values at the head of every derived column, with nothing raised anywhere. ## The real trade is comparability | threshold | leading edge of the output | what the early values mean | |---|---|---| | equal to the length | a block holding no value | every published value rests on the same number of observations | | one | full from the first position | the first values rest on one, then two, then three observations | | somewhere between | a shorter empty block | a declared floor on how few observations may stand behind a value | The point of the strict setting is not neatness. It is that **nothing in the output distinguishes a value computed from two observations from one computed from thirty**: same column, same type, same appearance on a chart, same treatment by whatever reads it next. A mean of two observations is far more variable than a mean of thirty, and a spread or a quantile over two observations carries almost no information at all — yet published in the same column it will be compared with the rest as though it were the same kind of number. ## What the strict setting costs A block of positions holding no value at the head of a column is not free either: - a later step that drops rows with any absent value removes those rows, and the data is now shorter than the input by an amount nobody wrote down; - if several derived columns use different lengths, the longest one silently decides how much of the head is lost for all of them; - a chart or an aggregate over the whole column now covers a shorter span than its title claims. So the choice is between publishing values you cannot compare and losing rows you did not intend to lose. Both are defensible; neither is a default worth inheriting silently. ## What to do 1. State the threshold explicitly, in the same place you state the length, so a reader sees both decisions together. 2. If you accept early values, carry the number of observations each answer used as a column beside it. A consumer can then filter on it instead of guessing, and a chart can grey out the region rather than pretending. 3. Decide once, at the boundary of the pipeline, whether the incomplete head is trimmed or kept — and if trimmed, record how many rows went and why. 4. Check the interior as well as the head. Count how many positions hold no value in the middle of the data; a non-zero count there tells you the threshold is also firing on holes or on quiet stretches, which is usually news.