How Much History Is Enough
How many periods a forecast needs to see a season, a trend or a driver effect, and what to do when the series is shorter than the question requires.
The amount of history a forecast needs depends entirely on what it is trying to see. Two full cycles to see a season, three to five years to distinguish a trend from a slow swing, and enough occurrences of a driver to tell its effect from noise.
The series discussed in “How Much History Is Enough” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for how employee monitoring works, explore the official product site can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
A series long enough for one of these is routinely used for all three, which is where most far-horizon error originates.
Additional context for “How Much History Is Enough” is available from monday.com work-management resources. External material is most useful when it prompts a specific validation check rather than replacing local judgement.
Seeing a weekly shape
Six to eight weeks is usually plenty, and more can be worse: the pattern changes, and averaging over a year dilutes the current shape with an old one.
- Weekly shape: six to eight weeks, and longer can be worse.
- Annual season: two complete years, three preferably.
- Trend: three to five years, with care.
- A driver's effect: enough occurrences to clear the noise.
This is the one case where a short window beats a long one.
Seeing an annual season
Two complete years minimum, three preferably. With one year there is nothing to distinguish a season from an event that happened once.
With two, a repeated August dip is probably seasonal. With three it can be stated with some confidence and its size estimated.
Seeing a trend
Three to five years, and even then with care. Over two years a slow cycle and a trend look identical, and the extrapolation goes in opposite directions depending on which it is.
Where the series is shorter, flat with a wide band is the honest forecast, and it is very rarely the one published.
Seeing a driver's effect
Enough occurrences to separate the effect from the variation around it. For a campaign that lifts volume by 20% against a daily noise of 10%, four or five campaigns is adequate. For one that lifts by 5%, twenty is not obviously enough.
The arithmetic is worth doing before claiming a driver works.
The series that is long but not usable
Common: six years of data, three definition changes, one system migration and a pandemic in the middle. The nominal length is six years; the usable length is whatever has elapsed since the last break.
Length should always be quoted as "usable since", not as "we have".
What to do with a short series
Three options, in order of preference. Borrow structure from a comparable series elsewhere in the organisation. Forecast at a coarser level where the series is longer and split down. Or forecast flat and say so.
All three are better than fitting a trend to eighteen months.
Borrowing a seasonal shape
Legitimate and underused. Where a new team has four months of history but sits inside an operation with five years, the operation's seasonal shape is usually a better estimate of the team's than nothing.
The assumption is stated: "seasonal shape taken from the parent operation; the team's own will replace it from next year."
Too much history
A real problem at the near horizon and an underrated one further out. Where the operation has genuinely changed — new channel, new product, new process — the older data describes something that no longer exists.
Weighting recent periods more heavily is the standard remedy, and it needs a reason rather than a default.
The pandemic years
Still a live question in most series. Those periods are not noise and they are not signal; they are a different regime, and including them distorts both the level and the seasonal shape.
The common treatments are to exclude them, to treat them as a labelled event, or to start the usable series afterwards. Whichever is chosen should be written down, because it changes the answer materially.
How much history to keep
More than the method needs, because the next method may need more and because the error record depends on it.
Storage is not the constraint. Storing the definitions alongside is, and that is a filing discipline rather than a technical one.
The question to ask first
Not "how much data do we have" but "what are we trying to see, and how many occurrences of it are in the data". The first question has a comfortable answer and the second has a useful one.
History that exists and cannot be reached
A separate problem from a short series: the data is in an archived system, or in a format nobody can read, or behind a retention policy that deleted it.
Finding this out early matters, because the recovery is a project with a lead time and it cannot be done in the week the forecast is needed.
The series that will be long enough later
Worth saying when a new operation is being set up: in two years this will support a seasonal estimate, and in four a trend. Until then the forecast is flat with a wide band and that is not a failure.
Saying it at the start prevents the annual argument about why the forecast is not improving.
Stating it next to the forecast
"Series usable from March 2022; two full seasons; trend not estimated." One line, and it tells the reader precisely how much weight the far end of the forecast can carry.
It also makes the case, without arguing, for keeping the series clean from now on.