Breaks: When the Past Stopped Applying
How to recognise that a series has changed regime rather than moved, what to do with the history before the break, and why waiting is usually right.
A structural break is the point at which the past stopped describing the present. After one, the history before it is not a weaker guide — it is a guide to a different operation, and using it produces a forecast of something that no longer exists.
The series discussed in “Breaks: When the Past Stopped Applying” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for employee monitoring data security, see the relevant operational guidance can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
Recognising one early is worth more than any modelling improvement, and recognising one too eagerly is a common and expensive mistake.
For an independent reference relevant to “Breaks: When the Past Stopped Applying”, consult McKinsey operations insights. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
What causes them
A new channel opening or closing. A pricing change. A major customer arriving or leaving. A process automation. A merger. A regulatory change that alters who contacts you and why.
- A new channel opening or closing; a pricing change.
- A major customer arriving or leaving.
- A process automated; a merger.
- A regulatory change altering who contacts you and why.
All of these are announced somewhere before they appear in the data, which is the practical point.
The announcement is the signal
Breaks are the one kind of change that can be anticipated rather than detected, because somebody in the organisation decided to cause them.
A planner who hears about the automation project in March has six months to prepare. One who detects it in the data in November has a forecast that has been wrong for two quarters.
Detecting one in the data
A level shift that persists, accompanied by a change in shape or variance. Unlike a definition change it usually ramps over weeks rather than stepping, and unlike an event it does not revert.
Three to four periods is the minimum before calling one, and calling it at two is how a long event gets mistaken for a new regime.
The cost of calling it too early
The history is discarded, the forecast is rebuilt on a handful of periods, the band widens enormously — and then the series returns to where it was.
This is why "wait one more period" is usually right, and why the decision is taken with the operations manager rather than alone.
What to do with the history before
Keep it, exclude it from fitting, and use it for one thing: the seasonal shape, which usually survives a break even when the level does not.
Borrowing the old season onto the new level is a reasonable first move and it should be labelled as borrowed.
Forecasting across a break you know is coming
Not from the series at all. From an estimate of the change, applied to the baseline: "automation removes an estimated 18% of category C contacts from June."
The estimate comes from whoever is making the change, it will be optimistic, and the forecast should say so and carry a band around it.
Why the estimate is always optimistic
Because it is made by the people whose project it is, usually before the detail is known, and usually without a view of displacement — the contacts that do not disappear but move to another channel.
A workable rule from practice is to carry the sponsor's figure and a half of it as two cases, and to say which is which.
The break that is several small ones
Harder and more common than a single dramatic change: a succession of small automations, each too small to see, together amounting to a slow regime change.
These show up as a trend and are usually treated as one. Treating them as what they are means the forecast can anticipate the next one rather than extrapolate the last.
Rebuilding after a break
Start the usable series at the break. Borrow the season. Forecast flat on the new level with a wide band until two or three periods are in. Narrow as evidence accumulates.
That progression is honest and it is much faster than it looks: within a quarter the band is usually back to something workable.
Telling the reader
In the header, in the same line as the horizon: "series restarted June 2025 after the self-service launch; seasonal shape carried over; trend not estimated."
It explains the wide band before anybody has to ask, and it stops the forecast being compared with pre-break figures as though they were the same measurement.
The break that only affects the shape
Worth distinguishing: a change that leaves the level alone and alters the distribution. Extended opening hours move demand within the day without changing the daily total.
A level-based check will see nothing, and the rostering forecast will be wrong every day until somebody re-estimates the profile.
Breaks in the driver rather than the demand
The relationship can break while the series looks continuous: the campaign still runs, the demand still arrives, and the link between them has halved.
This shows up in the remainder rather than in the level, which is one of the reasons to look at the remainder at all.
The review question
Once a quarter: has anything happened or been announced that would stop the history applying. Asked of the operations manager and of whoever runs change.
Two minutes, and it is the only early warning system a forecast has.