What it means
Any model that predicts one number from another leaves something over. If you predict monthly sales from advertising spend, actual sales will never land exactly on the prediction, and the difference is the error term for that month.
Statisticians distinguish between the true error, which is the unobservable difference in the underlying relationship, and the residual, which is the observed difference produced by the model you actually fitted. In everyday business use the two words are used more or less interchangeably.
The error term is not merely a nuisance to be minimised, because its behaviour tells you whether the model can be trusted. Errors should be random, centred on zero and roughly the same size across the range of the data, and any visible pattern in them is a sign that something important has been left out.
In finance this matters practically. When analysts run a regression of a share's returns against the market, the error term captures the company-specific component, which is exactly the part active fund managers claim to predict.
Common problems have specific names worth recognising. Heteroskedasticity means the errors get larger as the predictor grows, and autocorrelation means this month's error is related to last month's, and both mean the model's confidence intervals understate the true uncertainty.
In practice
Real-world examples.
Example
A financial analyst regressing a bank's share returns against a market index treats the error term as the bank-specific return. She studies its size across several years to judge how much of the share's movement comes from company news rather than from the direction of the market as a whole.
Example
A utility forecasting electricity demand from temperature finds that its errors are consistently large in winter and small in summer. That pattern reveals that the relationship between cold weather and consumption is not a straight line, and the modelling team adds a squared temperature term to correct it.
Example
A credit team building a default prediction model checks that residuals show no pattern across loan size before signing the model off. A pattern would mean the model is systematically wrong for large borrowers, which is precisely where the bank's biggest losses would come from.
Formula
Calculation
In a simple linear regression: Actual Y = a + bX + e, so the error term e = Actual Y - Predicted Y.
A retailer fits a regression of monthly sales against advertising spend and obtains the equation: Monthly sales = $12,000 + 4.5 x Advertising spend.
In March the business spent $20,000 on advertising.
Predicted sales: $12,000 + (4.5 x $20,000) = $12,000 + $90,000 = $102,000.
Actual sales in March: $109,500.
Error term: $109,500 - $102,000 = $7,500.
The model under-predicted by $7,500, or about 7% of the prediction. One such month tells you little, but if the errors are positive for six months in a row, something systematic is missing from the model, such as seasonality or a competitor's store closure.Case study
Seen in the real world.
This illustrative and fictional example concerns Bellamy Garden Centres, an invented chain of eight stores that built a simple model linking weekly sales to marketing spend. The fitted equation was Weekly sales = $40,000 + 3.2 x Marketing spend.
In one week the chain spent $15,000 on marketing, so the model predicted $40,000 + (3.2 x $15,000) = $88,000. Actual sales came in at $95,200, giving an error term of $7,200. The marketing manager treated this as proof that the campaign had over-delivered.
The finance director looked at the full year instead of a single week and found that every spring week showed a large positive error and every autumn week an equally large negative one. Averaged across the year the errors cancelled out, exactly as an ordinary least squares model forces them to, but they were plainly not random. The illustrative conclusion was that the model was missing a seasonal factor entirely, and that individual weekly errors said far more about the time of year than about any particular campaign. The chain rebuilt the model with a seasonal adjustment and a term for average weekly rainfall, after which the errors shrank sharply and stopped following a calendar pattern. Only then did the marketing manager have a defensible basis for claiming that a specific campaign had beaten expectations.
Watch out
Common mistakes.
- Treating a single large error as evidence that a decision worked, when errors are expected to vary around zero in any model.
- Judging a model only by how well it fits past data, without checking whether the errors show a pattern.
- Assuming a small error term proves causation, when a model can fit closely while the real driver sits outside it entirely.
Questions
People also ask.
What is the difference between an error term and a residual?
The error term is the unobservable true deviation in the underlying relationship, while the residual is the measured deviation from the model you actually estimated.
Should the error terms add up to zero?
In an ordinary least squares regression with an intercept the residuals sum to zero by construction, which is why their pattern matters more than their total.
Does a bigger error term always mean a worse model?
Not necessarily, because some outcomes are inherently noisy, so the fair test is whether the model beats simpler alternatives on data it has not seen.
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%Related