What is a forecast override? Judgment is an intervention in the data

A forecast meeting can end with one number on the slide and no record of the number that came before it. The system forecast is gone. The reason for the change is remembered as “market context.” The actual arrives later, but nobody can tell whether the judgment improved the call or merely changed it.

A forecast override is the recorded change from a baseline forecast to a final forecast at a declared cutoff. The change is an intervention in the data. It is not the outcome being forecast.

The forecast-value-added article owns the broader process comparison. The forecast-accuracy article owns the danger of a total error measure hiding offsetting patterns. This page owns the override record between baseline and final.

What does forecast override mean?

Keep the forecast objects separate:

The baseline is not necessarily “objective.” It is the preserved comparator. The final forecast is not necessarily “human.” A workflow, manager, algorithm, or committee can produce it. The evaluation needs the path and the comparison, not a story about who had better intuition.

How is an override measured?

At the declared unit, override magnitude is:

$\text{delta} = \text{final forecast} - \text{baseline forecast}$

For an actual value $A$, baseline error and final error can be written as:

$e_{\text{baseline}} = A - \text{baseline forecast}$

$e_{\text{final}} = A - \text{final forecast}$

Choose a loss function before evaluating the intervention. Under an absolute-error loss, a simple forecast-value-added comparison is:

$\text{FVA} = \lvert e_{\text{baseline}} \rvert - \lvert e_{\text{final}} \rvert$

Positive FVA means the final forecast has lower absolute error for that row. Under squared error, weighted error, service-level loss, or another decision loss, the number can change. A signed error also answers a different question from a magnitude of loss.

What does an override audit look like?

The six rows below are synthetic. They contain no company forecast, seller judgment, customer demand, or observed business result.

F-03 shows why an override can worsen the chosen loss. F-05 cannot be compared because the horizon changed. F-06 cannot be credited because the reason was recorded after the information boundary. The record preserves those failures instead of turning them into a story about forecast skill.

What do the studies contribute?

Fildes, Goodwin, Lawrence and Nikolopoulos studied judgmental adjustments in four supply-chain companies and reported variable performance across settings. Their findings include directional and size patterns, with upward and downward adjustments not behaving as one universal policy. Fildes, Goodwin and De Baets develop a forecast-value-added comparison across multiple organizations and business units and report variation rather than one rule for all contexts.

Those findings do not become a B2B pipeline benchmark. A SKU demand series with a statistical baseline is not the same object as an opportunity forecast. The transfer question needs a new design, a stable baseline, a declared horizon, and an actual outcome.

Which override fields should a team preserve?

• Timing: forecast cutoff and the moment the override was entered.
• Baseline: exact value and version that the judgment changed.
• Direction and size: signed delta in the same unit as the forecast.
• Reason: information, constraint, or assumption that motivated the change.
• Information set: what was known at the cutoff and what arrived later.
• Horizon and segment: the unit, time span, and population being forecast.
• Actual and loss: later observation, reconciliation status, and selected loss function.
• Disposition: improved, worsened, unchanged, incomparable, or not evaluable.

A missing reason does not prove the override was wrong. It makes the intervention harder to audit. A missing baseline makes the counterfactual comparison impossible.

How should a team review forecast overrides?

1. Freeze the forecast cutoff, unit, horizon, population, and loss function.
2. Preserve the baseline before any judgmental change.
3. Record override direction, magnitude, reason, information, and time.
4. Publish the final forecast as a separate version.
5. Join the same later actual to baseline and final values.
6. Calculate the chosen loss for both and retain incomparable or missing states.
7. Review patterns by direction, size, reason, horizon, segment, and owner or process.

The final forecast is a decision input. The actual is a later observation. The override is the intervention between them. Once those objects are separated, the team can ask whether judgment added value without assuming the answer.