What it means
A fitted model supplies a predicted value for each observation, and subtracting that fitted value from the observed outcome gives a residual. Positive and negative residuals can cancel in an average even when individual discrepancies are large.
Squaring the residuals prevents that cancellation when measuring their size, and adding the squares gives the residual sum of squares. Taking a suitable average and square root converts that squared-unit total into an estimated scale in the original units.
The divisor is not automatically the number of observations, because fitting parameters uses information from the sample. Under a usual full-rank linear model with n observations and p fitted parameters, residual degrees of freedom are n minus p.
Count the intercept when it is estimated. A simple straight-line model with an intercept and slope uses two parameters, so its conventional divisor is n minus two, while a larger model generally requires a different count rather than reusing that formula.
The distinction from the standard deviation of the original outcome also matters: original spread is measured around an average, whereas residual spread is measured around model predictions, so a useful model can account for variation that otherwise appears in the original spread. Compare models only with compatible outcomes, observations and estimation methods.
Multiplying every measured outcome by one thousand multiplies the scale by one thousand without changing the underlying relationship, so a smaller numerical figure from another unit is not evidence of improvement. Additional predictors can reduce the residual sum of squares on the fitted data while consuming degrees of freedom, so the adjusted scale can move differently from the raw squared-error total and more predictors do not automatically create a better model for new observations.
A roughly constant error variance supports interpreting one scale across the fitted range. If residuals widen sharply for larger customers, one pooled number can conceal different uncertainty, and time dependence and unusual observations also deserve separate checks.
The measure is not the uncertainty of every fitted coefficient, nor the standard error of an individual future prediction, as those calculations use further information about the model, inputs and estimation uncertainty. For a manager, request the units, sample size, parameter count and residual diagnostics beside the figure.
Compare the size with operational tolerances and test forecasts on relevant new data. A smaller historical scatter estimate is useful evidence, not a certificate of a sound decision.
In practice
Real-world examples.
Example
A fictional delivery-cost model has a residual standard deviation of $20. That describes unexplained scatter around its fitted costs, not a promise that every delivery will be within $20 of prediction. The manager checks the distribution and unusual cases.
Example
A report converts costs from dollars to thousands of dollars. The residual standard deviation changes from 2,000 to 2 without improving the model. Consistent units are essential when comparing reports.
Example
An analyst adds predictors and reduces historical squared errors. The team checks degrees of freedom and performance on new observations before keeping them.
Formula
Calculation
For the stated least-squares setting, residual standard deviation = sqrt(RSS / (n - p)), where RSS is the sum of squared residuals and p includes all fitted parameters. Require positive residual degrees of freedom.
With 12 observations, an estimated intercept and slope give p = 2. If RSS = 1,000 squared dollars, the scale is sqrt(1,000 / 10) = sqrt(100) = $10.
Using 12 instead of 10 would give sqrt(1,000 / 12), about $9.13, a different divisor that ignores the two fitted parameters. Neither result alone supplies an individual prediction interval.Case study
Seen in the real world.
Fictional case study: Cedar Logistics compares two depot-cost models using the same monthly observations. A presentation calls the model with the lower squared-error total automatically superior. The analyst also calculates residual standard deviations with the correct parameter counts, checks plots for changing spread and tests the models on later months.
Several added predictors improve the historical fit but do not help the later forecasts. Cedar retains a simpler model and documents its remaining scatter in dollars. Budget users can judge that scale against their tolerance while keeping a separate forecast uncertainty review.
Watch out
Common mistakes.
- Using n minus two for every regression. The appropriate parameter count depends on the model and includes an estimated intercept.
- Reading the residual scale as a guaranteed bound for every future observation. Prediction uncertainty requires additional calculations and assumptions.
- Comparing numbers across different units or datasets without adjustment. The outcome scale and observations affect the measure.
Questions
People also ask.
Is it the same as residual variance?
No. It is the square root of the variance estimate and therefore uses the outcome's original units rather than squared units.
Does a zero value prove perfect forecasting?
No. Zero fitted residuals describe the observed data under the fitted model. An overly flexible model can still fail on new data.
What should accompany the number?
State the units, fitted model, sample size, parameter count and assumptions. Include residual-pattern checks and relevant validation rather than presenting the scale alone.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%