What it means
Regression residuals are the differences between observed and fitted values, and in time-ordered data those differences may be related across adjacent observations. Durbin-Watson examines a pattern in those residuals rather than the level of the original variable alone.
Penn State's regression teaching material presents the test for autocorrelated errors and discusses remedial methods, using a sales regression example, and the important object is the model's error structure, not a universal trading signal. Positive serial correlation means neighbouring errors tend to move together, so a model may underpredict for several periods and then overpredict for several periods.
That persistence can indicate missing time dynamics or another specification problem. Negative serial correlation means adjacent errors tend to alternate, and the statistic should still be interpreted through a formal test rather than a visual rule alone.
The usual statistic lies between zero and four, and two is a useful reference point but not a universal pass mark. Sample size, regressors and the chosen significance level affect the decision thresholds.
Formal procedures can involve lower and upper critical bounds with an inconclusive region, and a manager should accept an inconclusive diagnostic rather than force the model into a clean pass or fail for a presentation. Data ordering matters, since applying an adjacent-observation test to randomly sorted records can change the interpretation.
The analysis should preserve the relevant time sequence and explain missing or irregularly spaced periods. Model assumptions also matter, because the conventional test has limitations including situations involving lagged dependent variables, and an analyst should use an appropriate alternative or extension instead of applying the same table mechanically.
Serial correlation can undermine conventional inference, as standard errors calculated under an inappropriate independence assumption may misstate uncertainty. A visually good fit or high explanatory power does not resolve that error-structure problem.
The next step is diagnosis, not automatic correction, because Penn State recommends considering omitted predictors and discusses methods for particular autoregressive error structures, whereas adding a variable just to move the statistic toward two can create a poorly justified model. The statistic examines first-order relationships, so more complex seasonal or higher-order dependence can require other diagnostics, and a result near two does not establish independence at every lag or show that forecast uncertainty has been fully assessed.
For a non-finance manager reviewing a forecast, ask whether residual patterns have been tested, and request the assumptions, test decision and effect on uncertainty. Do not convert a residual diagnostic into a confident statement about tomorrow's market return.
In practice
Real-world examples.
Example
A finance analyst fits monthly revenue to an industry indicator. Persistent positive residuals prompt a review of missing time dynamics rather than a declaration that the next month's revenue must rise. The analyst tests a seasonal term and checks whether the pattern weakens.
Example
A manager sees a Durbin-Watson value close to two. The analyst also checks seasonality and other diagnostics before describing the regression as adequate for planning. The manager is told that a value near two is one piece of evidence, not a certificate.
Example
A model includes the previous period's dependent variable. The statistician considers a suitable serial-correlation test rather than using conventional Durbin-Watson critical bounds without checking validity. The choice of test is recorded with the model documentation.
Formula
Calculation
Statistic d = sum of squared adjacent residual differences divided by the sum of squared residuals. For residuals 1, 1, 1 and 1, the numerator is zero and the denominator is four, giving d = 0. This tiny constructed sequence illustrates persistent residuals, not a valid inferential conclusion from a realistic sample.
For alternating residuals 1, -1, 1 and -1, the adjacent differences are -2, 2 and -2, whose squares are 4, 4 and 4, so the numerator is 12. The denominator is 1 + 1 + 1 + 1 = 4, giving d = 12 / 4 = 3, which is on the high side of two and reflects the alternating pattern.
For residuals 2, 1, -1 and -2, the differences are -1, -2 and -1, whose squares sum to 1 + 4 + 1 = 6, and the squared residuals sum to 4 + 1 + 1 + 4 = 10, so d = 6 / 10 = 0.6. The smooth drift gives a low value, again as an illustration only.Case study
Seen in the real world.
Fictional case: A manager interprets a low Durbin-Watson statistic in a sales model as proof that sales will keep increasing. The analyst explains that it concerns correlated forecast errors, then revises the specification and uncertainty analysis. Management updates the forecast review to examine residual behaviour without presenting the diagnostic as a directional market or sales prediction.
The fictional analyst adds a lagged driver and a seasonal term, refits the model and re-runs the diagnostic. The result moves closer to the reference value and falls outside the inconclusive region, so the team widens its stated forecast range to reflect the remaining uncertainty. The review pack now shows the test result next to the forecast so that readers see both.
Watch out
Common mistakes.
- Testing the raw price series and calling it a residual test.
- Treating two as a universal certification of model quality.
- Ignoring model limitations, inconclusive bounds or higher-order dependence.
Questions
People also ask.
Is it a stock-price forecast?
No. It diagnoses serial correlation in a regression's residuals.
Does a value near two prove independence?
No. Other lags, assumptions and diagnostics still matter.
Can a result be inconclusive?
Yes. Formal critical-bound procedures can leave an inconclusive region.
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%