What Good Looks Like at Each Horizon
There is no single standard for forecast accuracy. What error is achievable at each distance, where the floor comes from, and how to set a target that means something.
There is no universal standard for forecast accuracy, and any figure quoted as one — "forecasts should be within five per cent" — is describing somebody else's operation at an unstated horizon. What is achievable depends on the noise in your own series, the horizon, and how much of the demand you generate yourself.
The horizon in “What Good Looks Like at Each Horizon” should determine which current records are relevant and how much detail the forecast can honestly support. A planning team evaluating check this product explanation for employee monitoring at tech companies can connect recent time and project context with short-term capacity questions, while keeping longer-range assumptions in a separately governed model.
The useful standard is your own floor and your own baseline, both of which you can calculate.
For an independent reference relevant to “What Good Looks Like at Each Horizon”, consult IBM time-series overview. Compare the general material with the organisation’s actual horizon, data definitions and decision process.
The floor
The irreducible variation: how different two otherwise identical periods are. In a count process it is roughly the square root of the count, so a day averaging 100 has a floor around ten per cent and one averaging 10,000 has a floor around one.
No method goes below the floor. A day averaging 100 has a floor around ten per cent; one averaging 10,000 has a floor around one. The whole space for improvement lies between the naive baseline and that.
No method goes below it. Knowing it stops people demanding accuracy the process does not contain.
Calculating your own
Take comparable periods — same weekday, same position in month, no events, stable level — and measure how much they differ from each other. That spread is the floor.
Half an hour, and it is the most useful single number in the error discussion.
The baseline
What a naive method achieves: last comparable period, or a simple average. Every accuracy figure should be quoted against it.
The gap between the baseline and the floor is the whole space in which forecasting can add value, and it is often narrower than people assume.
Typical shapes by horizon
In a reasonably stable operation of moderate size, daily error at a fortnight might run six to ten per cent; weekly error at a quarter, ten to eighteen; monthly error at a year, fifteen to thirty.
Those are shapes, not standards. Yours will differ and the point is to know yours.
Why the far horizon is not three times worse
Because aggregation helps. A monthly figure averages out daily noise, so the far horizon gets a tailwind from its own coarseness even as the uncertainty about the world grows.
This is why comparing a daily error at two weeks with a monthly error at a year is meaningless.
Where the error should be concentrated
In the periods where something unusual happened, not spread evenly. A forecast with evenly distributed error across all periods is either operating at the floor or is missing something that affects everything.
Looking at the distribution of errors, not just the average, catches this.
Setting a target that means something
Not "95% accurate". Something like: "weekly error at twelve weeks below twelve per cent, against a naive baseline of seventeen and a floor of about seven, measured over the last twenty periods, with mean signed error within two per cent of zero."
Long, specific, and achievable or not in a way anybody can check.
When the target is met and the decision is still bad
Which happens, and is the signal that the error measure does not match the decision. Good on totals and poor on peaks, where the decision is about peaks.
The remedy is to change what is measured, not to tighten the target.
When to stop improving
When the error is close enough to the floor that the remaining gap is worth less than the effort to close it. That comparison needs the cost of a point of error, which is the subject of its own note here.
Most functions pass this point without noticing and keep improving the wrong horizon.
Improvement that is really aggregation
A warning: moving from daily to weekly reporting will improve every accuracy figure without improving anything. So will forecasting a larger unit.
Accuracy figures are only comparable at constant granularity and constant horizon, which is why both belong in the same line as the number.
Comparing with other organisations
Rarely useful and frequently requested. Published accuracy figures come from different series, at unstated horizons and granularities, with different amounts of self-generated demand.
The only legitimate use is a sanity check on something implausible. A target set from somebody else's figure is a target set from somebody else's noise.
Why size changes what is achievable
A large operation forecasts a proportionally smaller error than a small one on the same process, purely because the noise averages out.
Which means a small team being "worse" than a large one is usually arithmetic rather than competence, and the floor calculation is what demonstrates it.
What to report
Error, baseline, floor, horizon, granularity, periods. Six items, one line each, published with the forecast.
It takes a small table and it converts every subsequent conversation about whether forecasting is any good from an exchange of impressions into a reading of a record.