Bias, and the Correction That Fixes It
Once a forecast is shown to be biased, there are four possible fixes and only one of them is adding a number. How to choose, and how to record the choice.
A bias is a forecast that is wrong in the same direction repeatedly. Once it has been demonstrated — six or more same-signed periods out of eight — there are four possible responses, and reaching for the fudge factor is the last of them rather than the first.
The error measure in “Bias, and the Correction That Fixes It” needs both the original forecast and a trustworthy record of what actually happened. Teams researching automated timesheet workflows can use consult the official implementation page to strengthen the actual-work side of that comparison without treating activity data as proof of demand, performance or intent.
The order matters because three of the four fix the cause and one conceals it.
The assumptions behind “Bias, and the Correction That Fixes It” can also be tested against Databricks data and AI resources. Record which guidance applies to the available series and which does not.
The four responses
Find the missing driver and model it. Correct the series problem that is producing it. Change the method, where the bias is a known failure mode of the current one. Or apply a stated standing adjustment with a review date.
- Find the missing driver and model it.
- Correct the series problem producing it.
- Change the method, where the bias is its known failure mode.
- Apply a stated standing adjustment, with a review date.
In that order.
Looking for the missing driver
Start with the event calendar: is something regularly present that is not in it. Then the self-generated demand: dispatches, releases, backlog. Then the drivers already in the register, checking whether any has grown.
Most biases are found here, and the finding is usually something everybody in operations knew about.
Looking for the series problem
A definition change absorbed as a level. A set of days filled with zeros. A category that moved in or out. Each produces a persistent offset.
The change log, where one exists, resolves this in minutes.
Changing the method
Where the bias is a known failure of what is running. A moving average is biased during a trend, by roughly half the window's worth of change. A trend model is biased after the trend flattens.
Recognising that the bias is the method's signature failure is itself the diagnosis, and the fix is the method rather than an adjustment.
The standing adjustment
Last resort, and legitimate when the first three have been tried. The conditions: it is stated in the forecast, it has a size, a reason, a date and a review date, and it sits outside the model rather than inside it.
Outside the model matters: it keeps the underlying forecast honest so that when the cause is found, the correction can be removed cleanly.
The adjustment that becomes permanent
The failure mode. A +5% applied in 2023 for a reason nobody recorded, still running in 2026, now compensating for something that stopped happening.
A review date is the only defence, and it has to be a date rather than an intention.
Sizing the correction
The mean signed error over the biased periods, not the largest or the most recent. And less than that rather than more: correcting by two-thirds of the observed bias is safer, because part of what looks like bias is usually noise.
Over-correcting produces a bias in the opposite direction and a confused record.
Watching the correction
The signed error should move toward zero and stay there. If it overshoots, the correction was too large or the cause has changed.
Two or three periods tell you. The correction is not fire and forget.
Bias in only part of the forecast
Common: unbiased overall, under on peaks, over on quiet days. A single adjustment makes both worse.
This calls for a shape fix — usually the seasonal profile or the intraday distribution — rather than a level fix.
The bias that comes back
After a correction is removed, or seasonally. A bias that appears every autumn is not a bias; it is an unmodelled seasonal effect, and the fix is the seasonal component.
Checking whether the bias has a calendar before correcting it saves this mistake.
Recording it
Date, direction, size, periods observed, cause found or not, action taken, review date. Seven fields, one row.
Three years of that file is an institutional memory of how the forecast has actually behaved, and it is worth more than any individual correction in it.
Correcting at the wrong level
A level correction applied when the problem is in the shape makes the quiet periods worse while barely helping the peaks.
Before correcting, it is worth splitting the signed error by period type — peak, ordinary, quiet — because a bias that appears in only one of the three is a shape problem.
The correction that masks a break
The most expensive misdiagnosis. A structural change produces a persistent one-sided error, a standing adjustment is applied, and the break is never identified.
The forecast then tracks the new level by accident and will be wrong again the next time anything moves. Four same-signed periods should prompt a look at the change record before anything is added.
Saying it in the commentary
That the bias was found, what caused it, and what was done. In the commentary, once, plainly.
A forecast that announces its own correction is substantially more credible than one that quietly improves, and it makes the next correction easier to propose.