How this instrument works
When you fit a model — a regression line, a forecast, a trend curve — it produces a predicted value for every observation in your data. The residual is simply the gap between what actually happened and what the model said would happen: residual = observed − predicted. A positive residual means the model under-predicted (reality came in higher); a negative residual means it over-predicted (reality came in lower); a residual of exactly zero means the model nailed that particular point.
Residuals matter far beyond a single data point. Plot every residual in a dataset against its predicted value or against time, and patterns jump out that a single R² number would hide entirely: a random, structureless scatter around zero is what a well-fitting model should produce, while a curve, a funnel shape, or a cluster of large residuals in one region signals that the model is missing something — a nonlinear relationship, changing variance, or an outlier group it wasn't built to handle.
This calculator handles the single core computation — one observed value, one predicted value, one residual — which is the building block every larger regression-diagnostics workflow is built from, whether you're checking one suspicious data point by hand or writing the formula into a spreadsheet column that will run down an entire dataset.
- Enter the actual, real-world value into Observed (actual) value.
- Enter the value your model or fit produced for that same case into Predicted (fitted) value.
- Read Residual — this is observed minus predicted, so a positive number means the model under-predicted and a negative number means it over-predicted.
- Repeat for each data point if you're checking a full model; scanning the resulting residuals for patterns is more informative than looking at any single one in isolation.
Worked example — observed 82 against a model's predicted 78
Enter 82 into Observed (actual) value and 78 into Predicted (fitted) value. The instrument subtracts: 82 − 78.
Residual reads 4.000000 exactly. The model predicted 78 for this case, but the real, observed outcome was 82 — four units higher. Because the residual is positive, the model under-predicted here; if you were reviewing an entire dataset, you'd want to see residuals like this scattered both positive and negative with no obvious pattern, rather than every residual leaning the same direction.
Questions
Is a residual the same thing as an error?
They're related but not identical, and statisticians are careful about the distinction. The 'error' is the unobservable, true difference between an observation and the real underlying process that generated it — you can never know it exactly. A residual is what you actually compute: the difference between an observation and your model's estimate of it. A residual is an observable stand-in for the error, not the error itself.
What does a residual of zero mean?
It means the model's prediction exactly matched the observed value for that particular case — a perfect fit at that one point. It doesn't mean the model is perfect overall; a model can hit some points exactly while missing others by a wide margin, and a single zero residual tells you nothing about how the model performs elsewhere.
Why does the sign of the residual matter?
The sign tells you the direction of the miss, which is often as important as its size. A consistently positive residual across many points in one region of your data means the model is systematically under-predicting there; a consistently negative pattern means systematic over-prediction. Random small-and-large, positive-and-negative residuals scattered with no visible pattern is the signature of a model that's capturing the real relationship well.
How are residuals used to check if a model is any good?
By looking at their pattern, not just their size. Plotting residuals against predicted values or against an input variable should show a random scatter centered on zero if the model fits well. A visible curve suggests a missing nonlinear term; a residual spread that grows or shrinks systematically suggests non-constant variance; a cluster of unusually large residuals can flag outliers or a subgroup the model doesn't describe well.
Can residuals be used with any type of model, not just linear regression?
Yes — the observed-minus-predicted calculation applies to any model that outputs a predicted value for a given case, whether that's a straight-line regression, a polynomial curve, a time-series forecast, or a machine-learning model's output. The formula itself never changes; what changes is where the 'predicted' number comes from.