How to Measure Forecast Accuracy by Segment

A company-level accuracy number is an average of things that behave differently. Enterprise deals close on a different curve than mid-market deals. Renewals follow a different base rate than new business. Averaging them produces a figure that is technically correct and operationally useless, because it never tells you where to intervene.

Segment-level accuracy answers the question the blended number cannot: which part of the business is predictable and which part is guessing.

Why does a blended accuracy number hide the real problem?

Errors in opposite directions cancel, so a business with two broken segments can report excellent overall accuracy. A team that overcalls enterprise by $2 million and undercalls mid-market by $2 million lands the quarter on the number and learns nothing.

The cancellation is not rare. Enterprise forecasts tend to run optimistic because a small number of large deals carry the period, while renewals tend to run conservative because customer success teams prefer to be surprised upward. Those two biases point in opposite directions by design, and the blend reads clean.

What segments should you actually split on?

Split on the dimensions that change the shape of the close curve, not the ones that show up on the org chart. Deal size band and motion type are the two that matter most. Geography and industry usually add noise without adding signal. Start with two splits and add a third only when both are stable. A four-way split across two dimensions creates cells that hold too few deals to read, and unstable cells produce confident conclusions about noise.

How do you keep small segments from producing false conclusions?

Set a minimum cell size of roughly 30 closed deals per quarter, and move anything below it to a rolling four-quarter window. An enterprise segment closing eight deals a quarter cannot support a quarterly accuracy read. One deal slipping moves the number by more than any process change ever will.

Two guards keep small cells honest. Report dollar-weighted error alongside deal-count error, so a single large deal is visible as a single large deal rather than disguised as a systemic miss. And exclude any quarter where one opportunity exceeded 40 percent of the segment's forecast, then note the exclusion rather than hiding it.

What do you do with the differences once you see them?

Match the fix to the pattern, because segment error has three distinct causes and only one of them is coachable. The pattern in the data tells you which one you have.

Consistent error in one direction inside a single segment usually means the stage weights applied to that segment came from a different segment's history. Enterprise deals sitting in a mid-stage carry a lower in-quarter close probability than mid-market deals in the same stage, and a shared weight table charges both the same rate. Rebuilding weights per segment fixes it without touching a rep. Our note on weighted pipeline covers the rebuild.

Error that appears in every segment at once is not a segment problem. It usually means something changed in the market or in the business and the model is running on old assumptions. A new competitor pressuring price shows up as deal sizes shrinking across the board. Rising rates slowing buyer capital deployment shows up as win rates falling across the board. Neither is fixed by segmentation, though segmentation is how you confirm the cause is external.

Error concentrated in one team while the segment's other teams look clean is a process or judgment problem, and that is the case where coaching applies.

How does coverage interact with segment accuracy?

Coverage read at the segment level often contradicts the company number, and the segment read is the one to trust. Across ORM's customer base the typical company sits near 3.5x pipeline coverage, with individual companies as low as 1.4x and as high as 5x.

A company at a healthy 4x overall can hold that ratio while its enterprise segment sits at 1.8x and its mid-market segment sits at 6x. The blended figure says the quarter is covered. The segment split says one team has no path to its number and another is sitting on pipeline it cannot work. Neither fact is visible without the split, which is the same argument made in the 3x pipeline coverage rule is wrong.

Stale pipeline distorts this further and it does not distribute evenly. Around 10 percent or more of a typical pipeline has gone 12 months without a change in stage, close date, or amount, and it usually clusters in one segment. Strip it before computing coverage or the segment with the worst hygiene will look the best covered.

What should the segment accuracy report contain?

Signed error and absolute error per segment, dollar-weighted, at day one and at quarter close, with the cell size printed next to every figure. The cell size is what stops a reader from acting on eleven deals.

Signed error shows direction and identifies bias. Absolute error shows precision. Reporting only one of them produces the wrong action about half the time, since a segment alternating between plus 20 and minus 20 percent needs qualification work while a segment consistently at plus 8 percent needs recalibration.

Hold the definitions constant across segments. If one team counts a deal as closed at signature and another counts it at first invoice, the accuracy comparison measures accounting policy rather than forecasting skill. Definitions for the underlying measure sit in the forecast accuracy glossary entry.