Missing is not zero: what a gap in the data does to an average

An unknown is not a pass and not a zero.

A blank cell can mean several different things. The company may not publish the figure at all. It may publish it somewhere nobody has read yet. It may have published it for an older period only. In none of those cases can you assume the figure is zero, and in none of them has the company passed whatever test you were about to run.

This matters most when a spreadsheet or a screening tool meets a blank. Many tools quietly turn a blank into a zero or drop the row without saying so. Either choice changes the answer, and neither shows up in the result.

Count a gap as zero and the average halves.

Take six illustrative miners. Three publish a cost per bitcoin of 40,000, 50,000 and 60,000 dollars. The other three publish nothing. There are three common ways to average them, and they give two different answers.

Leave the gaps out and the average of the three published figures is 150,000 divided by 3, or 50,000 dollars. Count the gaps as zero and the same 150,000 is divided by 6, giving 25,000 dollars: a cost that no company in the set reports and that is far below every published figure. The lower number is not a finding about the industry. It is the result of a blank being read as a zero.

A filled-in gap looks exactly like data.

The third treatment fills each gap with the average of the figures you do have. The average stays at 50,000 dollars, so the result looks safe. But the table now shows six figures where only three were published, and nothing on the page says which three are real. The spread also looks tighter, because the three filled rows all sit exactly on the average.

Filling a gap with an estimate is sometimes reasonable, for example when a model needs a complete set of inputs. When it is done, the filled values should be marked as estimates, and any result should say how many of its inputs were published and how many were filled. A filled value is still an unknown.

A gap can also slip through a screen.

Suppose you screen for miners whose cost is below 45,000 dollars. With gaps counted as zero, the three silent miners pass, because zero is below every limit. With gaps kept as unknown, they neither pass nor fail: they need a figure before they can be judged.

Treated as zero, four of six pass, and three of the four passes rest on nothing. Kept as unknown, one passes, two fail and three are still open. The second answer is less tidy and more true.

A good table names each gap and why it is there.

The fix is to keep an unknown visible. Write words where the number would be, say why it is missing, and count how many inputs a total or an average actually used. Sort the unknowns to one end rather than mixing them in as zeros.

The real case behind the example: of the six miners in the cost article, three publish a cost per bitcoin with depreciation and three publish only a direct cost. The article's chart leaves a gap where the bars with depreciation of MARA and American Bitcoin would be, and does not draw IREN's direct figure, which is for the quarter ended Dec 31, 2025, and a second table lists those three miners with their direct cost only. Where a table has no figure for a company, such as IREN's revenue per bitcoin on the Companies page, it prints no revenue on file, and a click on a column sorts that row last. On a company page, a figure the site cannot show as a number prints words such as Not tagged or No SEC data in the number's place, and a filed value that fails a check against the previous period prints In review until it is checked. The coverage table on the methodology page names each gap with its reason.

Where these ideas come from.

The ideas in this article come from the free Clief Notes courses by Jake Van Clief, in particular David's guide to checking a public code repository before using it, in the David's Corner part of the courses, which treats missing evidence as an open question rather than a pass and an estimated middle value as unknown evidence. The courses are free at skool.com/cliefnotes. The examples and the wording here are MinerTerminal's own.