Root Mean Squared Error (RMSE) in Forecasting

RMSE scores a revenue forecast by squaring each period's error, averaging the squares, and taking the square root of that average. The squaring step is what separates RMSE from every other accuracy metric. Big misses count disproportionately, which matches how a revenue miss actually costs a company.

``` RMSE = square root of ( average of (Actual - Forecast)^2 ) ```

Why squaring changes the verdict

Two forecasts, four quarters each, identical average error in dollars.

Mean absolute error calls these forecasts equally good. RMSE says Forecast B is twice as bad. RMSE is right about the thing that matters operationally. Four quarters of steady $100K drift is a calibration problem you correct once. One quarter that misses by $400K is a hole in the plan that shows up as an emergency, and no amount of accuracy in the other three quarters buys it back.

RMSE never falls below MAE. When the two are close, errors are uniform. When RMSE runs far above MAE, the forecast is mostly fine and occasionally catastrophic, and the tail is where your risk lives.

RMSE stays in dollars

RMSE inherits the units of the thing being forecast, so a revenue RMSE is a dollar figure. That makes it directly readable against a decision. An RMSE of $600K on a quarterly forecast means a plan that commits every dollar of forecast revenue will be short by roughly that magnitude in a bad quarter.

The tradeoff is that dollar-denominated scores do not travel. You cannot compare an enterprise team's RMSE to an SMB team's RMSE, and you cannot compare this year's RMSE to last year's if the business doubled. Use RMSE inside a fixed scope and use a percentage metric across scopes.

What drives RMSE up

Timing variance is the usual culprit in B2B revenue, because a deal that closes one week past quarter end contributes its entire value to the error. ORM groups each opportunity with a machine learning model and predicts a close curve for the group, with those curves running from 1 to 80 weeks and most of the expectation landing before week 12. Groups with long, flat curves are the ones that generate the tail RMSE punishes.

The other driver is a model running on stale assumptions. ORM's read is that the most common reason a forecast fails is that something in the business or the market changed and the forecast was built on old assumptions. That failure does not produce steady drift. It produces one quarter that misses badly, which is exactly the shape RMSE is built to expose. Track it alongside forecast accuracy and your broader sales forecasting process metrics.