Trend, and the Line That Keeps Going
Fitting a trend is the commonest way to produce a large far-horizon error. What a real trend looks like, why trends flatten, and how damping fixes most of it.
A fitted trend extended four quarters forward is the commonest source of large far-horizon error in workforce forecasting. The arithmetic is sound; the assumption — that change continues at the rate it has been running — is nearly always wrong, because the things that cause trends in service demand are finite.
The series discussed in “Trend, and the Line That Keeps Going” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for small business time tracking, read more on this official resource can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
A trend is a statement about a cause. Where the cause is not named, the line should be treated with suspicion.
For an independent reference relevant to “Trend, and the Line That Keeps Going”, consult SciPy scientific-computing resources. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
What causes real trends
A migration between channels, which completes. A product ramping, which plateaus. A customer base growing, which can genuinely continue. A market being entered, which saturates.
- A migration between channels — completes.
- A product ramping — plateaus.
- A market being entered — saturates.
- A customer base growing — can genuinely continue.
Three of those four flatten, and the fourth is the only one where straight-line extrapolation is defensible.
Naming the cause
The practical test before fitting any trend: what is causing this, and is it finite? If the answer is "a self-service migration that is about sixty per cent complete", the forecast should flatten and the model should say so.
If there is no answer, the slope may be a slow cycle rather than a trend.
Trend versus slow cycle
Indistinguishable over two years and opposite in their implications. Three to five years of data usually separates them; less does not.
Where the series is short, the honest forecast is flat with a wide band, and that is very rarely what gets published.
Letting the slope decay
The single most useful correction. Instead of extending the slope at full strength, let it decay with distance so the forecast flattens over the horizon.
It costs nothing, it matches what trends actually do, and it removes most of the catastrophic far-horizon errors.
Choosing a damping rate
From your own history: fit a trend to each of several past periods, extend it, and see how far out it stopped being right. The typical answer is two to four periods.
That empirical answer is better than any default and takes an afternoon to produce.
The trend that is a step
A level shift fitted as a slope. The series jumped once, the fitted line reads it as continuous growth, and the forecast climbs forever.
This is the definition-change problem in its most expensive form, and plotting the series finds it immediately.
The trend in a ratio
Often more informative than the trend in the level. Contacts per customer, tasks per transaction, handling time per case. A level that is growing because the customer base is growing is a different fact from one growing because each customer contacts more.
The ratio trend is also usually more stable and more forecastable.
Trends that reverse
Channel migrations that stall. Growth that becomes decline. These are the cases where a trend model is worst, because it has momentum built in and reverses slowly.
A method that tracks the level quickly with damped trend handles reversal much better than one with a strong fitted slope.
Testing a trend out of sample
Hold back the last four periods, fit on the rest, and compare. Then do it again holding back a different four.
If the trend helps in some windows and hurts in others, it is not a trend; it is a cycle being fitted repeatedly in different directions.
What to publish when a trend exists
The level, the slope, the cause, and the damping. "Growing 1.8% a quarter, attributed to the migration from branch, damped to zero over four quarters because the migration is about two-thirds complete."
That sentence is auditable. A rising line is not.
Checking the trend against its cause
Quarterly, against the named cause. Has the migration completed, has the ramp plateaued, is the customer growth still happening.
The review is a conversation about the business and it takes minutes. It is also the only thing that will catch a flattening before the error record does, and by then two quarters have been planned wrongly.
The trend that is in the wrong units
A frequent confusion: growth in contacts and growth in workload are different if handling time is also moving. A series growing three per cent in volume and falling two per cent in handling time is flat in hours.
Since staffing is decided in hours, the hours trend is the one that matters, and it is rarely the one plotted.
The honest default
No trend unless one can be named and evidenced over at least three years, and damped when it is. That default is conservative, it loses a little in genuinely growing operations, and it avoids the error that costs most.