Direction: the Half Nobody Reads
A forecast that is wrong in the same direction every period is a correctable problem that mean absolute error cannot see. How to find it and what it usually means.
Eight quarters of one forecast, read by direction rather than by size
Under; billing peak moved.
Over; a launch that did not land.
Under; same billing effect.
Under.
Under.
Under.
Under.
Under.
Seven of the eight were under, by a mean of about six per cent, and the mean absolute error of 6.0% was reported as acceptable throughout. The size never changed; the direction never changed either, and only one of the two was ever looked at. One employer's own record, not a benchmark.
A forecast that is wrong by six per cent in a random direction and one that is wrong by six per cent in the same direction every period have the same absolute error and are completely different problems. The first is noise and cannot be improved. The second is bias, it is correctable, and it is invisible in every accuracy measure in common use.
The method in “Direction: the Half Nobody Reads” is only as useful as the definitions and source records behind it. A team exploring a transparent time tracking with screenshots process for time tracking with screenshots can add timely work and project context, with transparent settings, limited access and a documented correction route around every consequential use.
Reading direction is a column of plus and minus signs. It takes a minute and it is almost never done.
Teams documenting “Direction: the Half Nobody Reads” may use Statsmodels documentation as an outside reference. Test the method against the real forecast horizon, known events and operational constraints.
What bias looks like
A run of same-signed errors. Six of eight under, or seven of eight over. The individual misses look unremarkable; the pattern does not.
The test is simply counting signs. Seven of eight in one direction is unlikely enough to act on; five of eight is not.
Why it hides
Because mean absolute error discards the sign by construction, and because each individual period has an explanation. The billing peak moved. The campaign was bigger than expected. Each is plausible on its own.
Only the sequence shows that the same thing is happening every time.
The four usual causes
A driver whose size has been underestimated. A missing driver that is always present. A definition change absorbed as a level. Or a target influencing the forecast.
Each has a different fix, and all four are findable once the direction is known.
The under-forecast that is a missing driver
The commonest. Something adds demand regularly and is not in the model — a billing effect, a standing displacement from another channel, self-generated contacts from the operation's own backlog.
Because it is always present, it looks like the forecast being a bit low rather than like a missing term.
Correcting a bias
Not by adding a fudge. By finding the cause and modelling it. A bias corrected by a constant adjustment will hold until the cause changes, and then it will be wrong by twice as much.
Where the cause genuinely cannot be found, a stated correction with a review date is acceptable as an interim.
The honest interim
"A standing adjustment of +5% is applied from March, because the forecast has been under in seven of the last eight quarters and the cause has not been identified. To be reviewed in September."
That sentence is defensible. An unlabelled +5% inside the model is not.
Bias that is deliberate
Legitimate where the cost of being under differs from the cost of being over, which it usually does. Staffing at the 70th percentile is a sensible policy.
The condition is that it lives in the staffing rule and not inside the forecast, so that the forecast remains a statement about what is expected.
Bias by horizon
Often different at different distances. A forecast unbiased at two weeks and six per cent under at twelve is telling you something specific: the near method is fine and the trend or the long-horizon driver is wrong.
Separating the record by horizon is what makes this visible.
Bias by period type
Also worth splitting: biased on peaks but not on ordinary days, biased on Mondays, biased in the months containing a particular event.
Each split that shows a difference is a diagnosis. Splits that show nothing are also useful, because they eliminate causes.
How long before acting
Four periods for a signal, six to eight before correcting. Acting on three is reacting to noise; waiting for twelve is a year of decisions taken on a known error.
The chart
Signed error by period, with a zero line. Eight bars. Anybody can read it in two seconds and it shows the thing that no table of percentages shows.
It is the single most useful chart a forecasting function can put in its pack, and it is almost never in one.
Direction at the level of components
Worth splitting once a bias is confirmed. The level may be unbiased while the seasonal component runs low, or the baseline fine and the event adjustments consistently short.
Each component's own signed error points at a different fix, and the combined figure points at none of them.
Bias that the organisation prefers
Occasionally a persistent direction is deliberate and nobody wrote it down — a planner quietly forecasting high because being short is punished.
That is a rational response to an incentive and it belongs in the staffing rule where it can be discussed, not in the forecast where it cannot.
Why this matters more than accuracy
Because bias is the part that can be fixed. The random part of the error is the world; the systematic part is a mistake, and it is sitting in plain view in a column of signs nobody has added up.