# What Is the Durbin-Watson Test?

Published: 2026-02-23
Author: Warren Team
URL: https://www.heywarren.com/blog/durbin-watson-test

---
Every year, analysts publish financial forecasts built on regression models that quietly violate a foundational assumption, and almost nobody checks. The result is confidence intervals that lie, predictions that drift, and investment decisions made on flawed math. The durbin watson test exists precisely to catch this hidden flaw before it does damage.

Most people who run a regression never test for serial correlation. They check R-squared, glance at p-values, and call it done. But if the residuals in their model are correlated with each other across time, a problem called autocorrelation, every standard error, every t-statistic, and every confidence interval they calculated is wrong. Not slightly off. Systematically wrong.

By the end of this guide, you will understand what the Durbin-Watson test measures, how to interpret its output, and why it matters for anyone building financial models, evaluating economic data, or making evidence-based investment decisions. You will also learn the most common mistakes analysts make and what to do when the test reveals a problem.

The Durbin-Watson statistic has been a cornerstone of econometrics since James Durbin and Geoffrey Watson introduced it in 1950. Decades later, it remains one of the most widely used diagnostic tools in financial and economic research.

---

## What Is the Durbin-Watson Test?

The Durbin-Watson test is a statistical diagnostic tool that checks whether the residuals, the errors, from a regression model are correlated with each other across sequential observations. It produces a single number, called the Durbin-Watson statistic, that ranges from 0 to 4. A value near 2 means no autocorrelation. Values below 2 signal positive autocorrelation. Values above 2 signal negative autocorrelation.

In plain English: if yesterday's forecasting error predicts today's forecasting error, your model has autocorrelation. That is exactly the pattern the Durbin-Watson test is designed to detect.

The test applies specifically to **ordinary least squares (OLS) regression**, the most common type of regression used in finance and economics. It assumes you are working with data ordered in time, quarterly earnings, monthly returns, daily interest rates, where observations have a natural sequence.

**Autocorrelation** (also called **serial correlation**) violates one of the core assumptions of OLS regression. When that assumption breaks down, the coefficient estimates are still unbiased, but the standard errors are no longer reliable. Hypothesis tests produce misleading results, and confidence intervals become too narrow, making your model look more precise than it actually is.

---

## How the Durbin-Watson Test Works

Understanding the mechanics behind the test helps you use it correctly and interpret results with confidence.

![The four-step process for computing the Durbin-Watson statistic from an OLS regression.](data:image/svg+xml,%3Csvg%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%20viewBox%3D%220%200%20800%20149%22%20width%3D%22800%22%20height%3D%22149%22%20role%3D%22img%22%3E%3Ctitle%3ETimeline%3C%2Ftitle%3E%3Crect%20width%3D%22100%25%22%20height%3D%22100%25%22%20fill%3D%22%23f8fafc%22%2F%3E%3Cline%20x1%3D%22137.5%22%20y1%3D%2255%22%20x2%3D%22662.5%22%20y2%3D%2255%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%223%22%2F%3E%3Ccircle%20cx%3D%22137.5%22%20cy%3D%2255%22%20r%3D%2224%22%20fill%3D%22white%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22137.5%22%20y%3D%2260%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2215%22%20font-weight%3D%22700%22%20fill%3D%22%230f172a%22%3E1%3C%2Ftext%3E%3Ctext%20x%3D%22137.5%22%20y%3D%22101%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2212%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3ERun%20OLS%3C%2Ftext%3E%3Ctext%20x%3D%22137.5%22%20y%3D%22119%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3EFit%20regression%20model%3C%2Ftext%3E%3Ccircle%20cx%3D%22312.5%22%20cy%3D%2255%22%20r%3D%2224%22%20fill%3D%22white%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22312.5%22%20y%3D%2260%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2215%22%20font-weight%3D%22700%22%20fill%3D%22%230f172a%22%3E2%3C%2Ftext%3E%3Ctext%20x%3D%22312.5%22%20y%3D%22101%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2212%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EGet%20Residuals%3C%2Ftext%3E%3Ctext%20x%3D%22312.5%22%20y%3D%22119%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3EObserved%20minus%20predicted%3C%2Ftext%3E%3Ccircle%20cx%3D%22487.5%22%20cy%3D%2255%22%20r%3D%2224%22%20fill%3D%22white%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22487.5%22%20y%3D%2260%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2215%22%20font-weight%3D%22700%22%20fill%3D%22%230f172a%22%3E3%3C%2Ftext%3E%3Ctext%20x%3D%22487.5%22%20y%3D%22101%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2212%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EDiff%20Squared%3C%2Ftext%3E%3Ctext%20x%3D%22487.5%22%20y%3D%22119%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3E%CE%A3%28e%E2%82%9C%20%E2%88%92%20e%E2%82%9C%E2%82%8B%E2%82%81%29%C2%B2%3C%2Ftext%3E%3Ccircle%20cx%3D%22662.5%22%20cy%3D%2255%22%20r%3D%2224%22%20fill%3D%22%232563eb%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%223%22%2F%3E%3Ctext%20x%3D%22662.5%22%20y%3D%2260%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2215%22%20font-weight%3D%22700%22%20fill%3D%22white%22%3E4%3C%2Ftext%3E%3Ctext%20x%3D%22662.5%22%20y%3D%22101%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2212%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EDivide%3C%2Ftext%3E%3Ctext%20x%3D%22662.5%22%20y%3D%22119%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3EBy%20%CE%A3e%E2%82%9C%C2%B2%3C%2Ftext%3E%3C%2Fsvg%3E)

*The four-step process for computing the Durbin-Watson statistic from an OLS regression.*

### The Formula Behind the Statistic

The Durbin-Watson statistic (abbreviated **d**) is calculated using the following logic:

1. Run your OLS regression and collect the **residuals**, the differences between observed values and predicted values.
2. Calculate the sum of squared differences between consecutive residuals.
3. Divide that sum by the total sum of squared residuals.

More formally: d = Σ(eₜ − eₜ₋₁)² / Σeₜ²

Where eₜ is the residual at time t and eₜ₋₁ is the residual from the prior period. If consecutive residuals tend to move together (positive autocorrelation), the numerator will be small relative to the denominator, pushing d toward 0. If they tend to alternate in sign (negative autocorrelation), d moves toward 4.

### The Role of the Null Hypothesis

The Durbin-Watson test frames autocorrelation detection as a hypothesis test. The **null hypothesis** is that no first-order serial correlation exists in the residuals. The **alternative hypothesis** is that positive or negative autocorrelation is present.

Unlike many statistical tests, the Durbin-Watson test does not use a single critical value. It uses two boundary values, dL (lower) and dU (upper), that depend on your sample size and the number of predictors in your model. These critical values come from published tables.

- If d < dL: reject the null hypothesis; positive autocorrelation is present.
- If d > dU and d < (4 − dU): fail to reject the null hypothesis; no significant autocorrelation.
- If dL ≤ d ≤ dU or (4 − dU) ≤ d ≤ (4 − dL): the test is **inconclusive**.

This inconclusive zone is one of the test's known quirks, and it is why analysts sometimes supplement it with other methods.

---

## Interpreting Your Durbin-Watson Statistic

The number your software returns only becomes useful once you know how to read it in context. Here is a practical breakdown.

![The 0–4 Durbin-Watson scale and what each range signals about autocorrelation in your model.](data:image/svg+xml,%3Csvg%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%20viewBox%3D%220%200%20600%20211%22%20width%3D%22600%22%20height%3D%22211%22%20role%3D%22img%22%3E%3Ctitle%3EHierarchy%3C%2Ftitle%3E%3Crect%20width%3D%22100%25%22%20height%3D%22100%25%22%20fill%3D%22%23f8fafc%22%2F%3E%3Crect%20x%3D%22220%22%20y%3D%2220%22%20width%3D%22160%22%20height%3D%2258%22%20rx%3D%228%22%20fill%3D%22%232563eb%22%2F%3E%3Ctext%20x%3D%22300%22%20y%3D%2254%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22700%22%20fill%3D%22white%22%3EDW%20Statistic%20%280%E2%80%934%29%3C%2Ftext%3E%3Cpath%20d%3D%22M%20300%2078%20L%20300%20105.5%20L%20120%20105.5%20L%20120%20133%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%20fill%3D%22none%22%2F%3E%3Crect%20x%3D%2240%22%20y%3D%22133%22%20width%3D%22160%22%20height%3D%2258%22%20rx%3D%228%22%20fill%3D%22white%22%20stroke%3D%22%230891b2%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22120%22%20y%3D%22158%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2213%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3E0%20%E2%80%93%202%3C%2Ftext%3E%3Ctext%20x%3D%22120%22%20y%3D%22176%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3EPositive%20autocorr.%3C%2Ftext%3E%3Cpath%20d%3D%22M%20300%2078%20L%20300%20105.5%20L%20300%20105.5%20L%20300%20133%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%20fill%3D%22none%22%2F%3E%3Crect%20x%3D%22220%22%20y%3D%22133%22%20width%3D%22160%22%20height%3D%2258%22%20rx%3D%228%22%20fill%3D%22white%22%20stroke%3D%22%230891b2%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22300%22%20y%3D%22158%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2213%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3E%E2%89%88%202%3C%2Ftext%3E%3Ctext%20x%3D%22300%22%20y%3D%22176%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3ENo%20autocorr.%3C%2Ftext%3E%3Cpath%20d%3D%22M%20300%2078%20L%20300%20105.5%20L%20480%20105.5%20L%20480%20133%22%20stroke%3D%22%23cbd5e1%22%20stroke-width%3D%222%22%20fill%3D%22none%22%2F%3E%3Crect%20x%3D%22400%22%20y%3D%22133%22%20width%3D%22160%22%20height%3D%2258%22%20rx%3D%228%22%20fill%3D%22white%22%20stroke%3D%22%230891b2%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22480%22%20y%3D%22158%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2213%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3E2%20%E2%80%93%204%3C%2Ftext%3E%3Ctext%20x%3D%22480%22%20y%3D%22176%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2210%22%20fill%3D%22%2364748b%22%3ENegative%20autocorr.%3C%2Ftext%3E%3C%2Fsvg%3E)

*The 0–4 Durbin-Watson scale and what each range signals about autocorrelation in your model.*

### The 0-to-4 Scale in Plain Terms

The full range of possible Durbin-Watson values maps onto three distinct outcomes:

- **0 to less than 2**: Positive autocorrelation. Residuals trend in the same direction across time. This is the most common problem in financial time series data. Stock returns, GDP growth rates, and unemployment figures often exhibit this pattern.
- **Exactly 2**: The ideal result. No autocorrelation detected. Consecutive residuals are uncorrelated.
- **Greater than 2 to 4**: Negative autocorrelation. Residuals alternate signs, a positive error is followed by a negative error, then positive again. This is less common but still problematic.

In practice, most analysts treat any value between 1.5 and 2.5 as acceptable and flag anything outside that range for further investigation.

### Reading Results Against Critical Value Tables

Software like R, Python (statsmodels), Stata, and Excel can calculate the Durbin-Watson statistic automatically. But interpreting it requires checking against critical values from the Durbin-Watson distribution tables.

For example, suppose you have 50 observations and 3 predictor variables. The critical values at the 5% significance level are approximately dL = 1.42 and dU = 1.67.

- d = 1.20 → below dL → positive autocorrelation confirmed.
- d = 1.55 → between dL and dU → inconclusive.
- d = 1.90 → above dU → no significant autocorrelation.

Most modern software packages report a p-value alongside the statistic, which removes the need to look up tables manually.

---

## Why Autocorrelation Matters in Financial Modeling

Serial correlation is not an abstract mathematical concern. It has real consequences for investors, analysts, and anyone relying on quantitative models to make decisions.

When a **time series regression** suffers from autocorrelation, the estimated standard errors of the coefficients are biased, typically downward. This makes t-statistics larger than they should be. A variable that appears statistically significant at the 5% level might fail significance once autocorrelation is corrected. Analysts who skip the Durbin-Watson test risk building their entire strategy on a false positive.

Consider a portfolio manager who regresses monthly stock returns against three economic indicators. If the model's residuals are positively autocorrelated, meaning last month's prediction error predicts this month's, the manager might believe the model explains 65% of return variance when the true figure is closer to 40%. The model looks precise. The allocation decisions feel data-driven. But the underlying math is compromised.

**Three specific financial contexts where autocorrelation is especially dangerous:**

- **Factor models**: If residuals in a Fama-French three-factor regression are autocorrelated, risk-adjusted return estimates are unreliable.
- **Macro forecasting**: Economic variables like inflation and unemployment are highly persistent. Regressions using them as predictors frequently produce autocorrelated residuals.
- **Algorithmic trading strategies**: A trading signal built on a misspecified regression may appear profitable in backtests but fail in live markets because the statistical foundation is flawed.

The durbin watson test is the first line of defense against these errors. Running it takes 30 seconds. Ignoring it can cost far more.

---

## The Durbin-Watson Test in Practice: Real-World Examples

Abstract theory is easier to absorb when grounded in concrete scenarios. Here are two examples that illustrate how the test functions in applied financial work.

### Example 1: Housing Price Regression

An economist models median home prices in 80 U.S. metropolitan areas over 10 years using mortgage rates, income levels, and population growth as predictors. After running OLS regression, they calculate a Durbin-Watson statistic of 0.87.

With 800 observations and 3 predictors, the lower critical value at 5% significance is approximately dL = 1.72. Since 0.87 < 1.72, the test rejects the null hypothesis. Strong positive autocorrelation is present, not surprising given that housing prices trend persistently over time.

The economist applies the **Cochrane-Orcutt procedure**, a method that iteratively estimates and corrects for autocorrelation. After correction, the Durbin-Watson statistic rises to 1.91, falling comfortably within the acceptable range. The corrected model produces wider, more honest confidence intervals.

### Example 2: Earnings Forecast Model

An analyst at a regional bank builds a quarterly earnings forecast model for a retail company, using consumer spending, same-store sales growth, and credit card delinquency rates as inputs. The initial Durbin-Watson result is 2.3.

With 40 quarters of data and 3 predictors, the upper critical value is approximately dU = 1.66. Since 2.3 is above (4 − 1.66) = 2.34... wait, 2.3 is actually less than 2.34, so it falls in the no-autocorrelation zone. The analyst proceeds with the original model, confident that serial correlation is not inflating the apparent precision of the forecasts.

These two examples show how the same test can produce very different action steps depending on the result.

---

## Common Mistakes When Using the Durbin-Watson Test

Even analysts who know the Durbin-Watson test exists often apply it incorrectly. Avoiding these pitfalls will save time and prevent misinterpretation.

**Mistake 1: Applying it to non-time-ordered data.**
The Durbin-Watson test is designed for data with a natural time sequence. If your observations are cross-sectional, say, 200 companies at a single point in time, the test is not appropriate. Use spatial autocorrelation tests or other diagnostics instead.

**Mistake 2: Treating the inconclusive zone as "no problem."**
If your d value falls between dL and dU, the test is genuinely inconclusive. That does not mean autocorrelation is absent. Run supplementary tests like the **Breusch-Godfrey test**, which handles more complex autocorrelation patterns and has no inconclusive zone.

**Mistake 3: Only checking for first-order autocorrelation.**
The standard Durbin-Watson test only detects **first-order serial correlation**, correlation between adjacent observations. It does not catch higher-order autocorrelation, where residuals correlate across two, three, or more periods. Annual data with quarterly patterns may fool a basic Durbin-Watson check.

**Mistake 4: Using it with lagged dependent variables.**
If your regression includes a lagged version of the dependent variable as a predictor, common in dynamic economic models, the Durbin-Watson test is biased. In that case, the Breusch-Godfrey test is the appropriate alternative.

**Mistake 5: Fixing autocorrelation with the wrong remedy.**
Autocorrelation in residuals sometimes signals **model misspecification**, a missing variable or the wrong functional form, rather than a pure serial correlation problem. Adding a lag of the dependent variable, including an omitted predictor, or transforming variables to percentage changes often resolves the issue more cleanly than statistical corrections alone.

---

## Limitations and Alternatives to the Durbin-Watson Test

No single test captures every dimension of model adequacy. Understanding where the Durbin-Watson test falls short helps you choose the right tools for each situation.

![Recommended decision flow for applying the Durbin-Watson test and choosing a follow-up action.](data:image/svg+xml,%3Csvg%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%20viewBox%3D%220%200%201090%20125%22%20width%3D%221090%22%20height%3D%22125%22%20role%3D%22img%22%3E%3Ctitle%3EFlow%20diagram%3C%2Ftitle%3E%3Crect%20width%3D%22100%25%22%20height%3D%22100%25%22%20fill%3D%22%23f8fafc%22%2F%3E%3Crect%20x%3D%2230%22%20y%3D%2225%22%20width%3D%22170%22%20height%3D%2275%22%20rx%3D%2210%22%20fill%3D%22white%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22115%22%20y%3D%2267.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3ERun%20DW%20Test%3C%2Ftext%3E%3Cline%20x1%3D%22205%22%20y1%3D%2262.5%22%20x2%3D%22237%22%20y2%3D%2262.5%22%20stroke%3D%22%2364748b%22%20stroke-width%3D%222%22%2F%3E%3Cpolygon%20points%3D%22244%2C62.5%20235%2C57.5%20235%2C67.5%22%20fill%3D%22%2364748b%22%2F%3E%3Crect%20x%3D%22245%22%20y%3D%2225%22%20width%3D%22170%22%20height%3D%2275%22%20rx%3D%2210%22%20fill%3D%22white%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22330%22%20y%3D%2267.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3ECheck%20d%20Value%3C%2Ftext%3E%3Cline%20x1%3D%22420%22%20y1%3D%2262.5%22%20x2%3D%22452%22%20y2%3D%2262.5%22%20stroke%3D%22%2364748b%22%20stroke-width%3D%222%22%2F%3E%3Cpolygon%20points%3D%22459%2C62.5%20450%2C57.5%20450%2C67.5%22%20fill%3D%22%2364748b%22%2F%3E%3Crect%20x%3D%22460%22%20y%3D%2225%22%20width%3D%22170%22%20height%3D%2275%22%20rx%3D%2210%22%20fill%3D%22white%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22545%22%20y%3D%2258.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EInconclusive%3F%3C%2Ftext%3E%3Ctext%20x%3D%22545%22%20y%3D%2278.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2211%22%20fill%3D%22%2364748b%22%3ERun%20Breusch-Godfrey%3C%2Ftext%3E%3Cline%20x1%3D%22635%22%20y1%3D%2262.5%22%20x2%3D%22667%22%20y2%3D%2262.5%22%20stroke%3D%22%2364748b%22%20stroke-width%3D%222%22%2F%3E%3Cpolygon%20points%3D%22674%2C62.5%20665%2C57.5%20665%2C67.5%22%20fill%3D%22%2364748b%22%2F%3E%3Crect%20x%3D%22675%22%20y%3D%2225%22%20width%3D%22170%22%20height%3D%2275%22%20rx%3D%2210%22%20fill%3D%22white%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22760%22%20y%3D%2258.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EAutocorr.%20Found%3F%3C%2Ftext%3E%3Ctext%20x%3D%22760%22%20y%3D%2278.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2211%22%20fill%3D%22%2364748b%22%3EFix%20or%20respecify%3C%2Ftext%3E%3Cline%20x1%3D%22850%22%20y1%3D%2262.5%22%20x2%3D%22882%22%20y2%3D%2262.5%22%20stroke%3D%22%2364748b%22%20stroke-width%3D%222%22%2F%3E%3Cpolygon%20points%3D%22889%2C62.5%20880%2C57.5%20880%2C67.5%22%20fill%3D%22%2364748b%22%2F%3E%3Crect%20x%3D%22890%22%20y%3D%2225%22%20width%3D%22170%22%20height%3D%2275%22%20rx%3D%2210%22%20fill%3D%22white%22%20stroke%3D%22%232563eb%22%20stroke-width%3D%222%22%2F%3E%3Ctext%20x%3D%22975%22%20y%3D%2267.5%22%20text-anchor%3D%22middle%22%20font-family%3D%22system-ui%2C-apple-system%2Csans-serif%22%20font-size%3D%2214%22%20font-weight%3D%22600%22%20fill%3D%22%230f172a%22%3EModel%20Valid%3C%2Ftext%3E%3C%2Fsvg%3E)

*Recommended decision flow for applying the Durbin-Watson test and choosing a follow-up action.*

**Key limitations:**

- It only tests for **first-order autocorrelation**. Higher-order patterns require the Breusch-Godfrey test or the Ljung-Box test.
- It performs poorly with **small samples** (fewer than 15 observations). Critical value tables become unreliable at that scale.
- It cannot be used when the model includes **lagged dependent variables** as regressors.
- The **inconclusive zone** creates ambiguity that the test cannot resolve on its own.

**Practical alternatives and complements:**

- **Breusch-Godfrey test**: More flexible; handles higher-order autocorrelation and lagged regressors. This is the preferred test when Durbin-Watson falls short.
- **Ljung-Box test**: Widely used in time series analysis; tests multiple lag orders simultaneously. Common in ARIMA model diagnostics.
- **ACF plots (autocorrelation function)**: Visual tool that shows the correlation between residuals and their lags. A spike at lag 1 suggests first-order autocorrelation; spikes at multiple lags suggest higher-order patterns.
- **Newey-West standard errors**: If autocorrelation is detected but the focus is on correcting standard errors rather than re-estimating the model, Newey-West robust standard errors are a widely accepted fix that requires no model transformation.

For most financial modeling work, running the Durbin-Watson test first, then confirming with a Breusch-Godfrey test if results are borderline, gives you a robust diagnostic workflow.

---

## Related Reading

**More from Warren**:
- [GMV: What Gross Merchandise Value Means and How to Calculate It](/blog/what-is-gmv)
- [What Is a Feasibility Analysis? TELOS Framework Guide](/blog/feasibility-analysis)
- [WACC Formula: How to Calculate Weighted Average Cost of Capital](/blog/wacc-formula)
- [Fixed Rate Swap: How Interest Rate Swaps Work and Who Uses Them](/blog/fixed-rate-swap)
- [Why's Everything Made in China: The Direct Answer](/blog/whys-everything-made-in-china)
- [What Do Controllers Do? The Core Role Explained](/blog/what-do-controllers-do)

## Authoritative Sources

For deeper background and primary-source data on this topic, the following authoritative sources are useful starting points:

- [IRS](https://www.irs.gov/)
- [SEC](https://www.sec.gov/)
- [Federal Reserve](https://www.federalreserve.gov/)
- [Consumer Financial Protection Bureau](https://www.consumerfinance.gov/)
- [U.S. Department of the Treasury](https://home.treasury.gov/)
- [Bureau of Labor Statistics](https://www.bls.gov/)

## Conclusion

The durbin watson test is one of the most practical tools in quantitative finance and econometrics, and one of the most frequently skipped. Here is what matters most:

- **The Durbin-Watson statistic ranges from 0 to 4.** Values near 2 indicate no autocorrelation. Values significantly below 2 signal positive serial correlation; values significantly above 2 signal negative serial correlation.
- **Autocorrelation does not bias your coefficient estimates, but it does distort your standard errors**, making models appear more precise than they are and producing misleading hypothesis test results.
- **The test has real limitations.** It only catches first-order autocorrelation and cannot be used with lagged dependent variables. Supplement it with the Breusch-Godfrey test when needed.
- **When autocorrelation is detected, the first question is whether the model is misspecified**, not just whether a statistical correction is needed. A missing variable often explains the pattern.
- **Modern software makes the test trivial to run.** There is no good reason to skip it when working with time series regression data.

Financial models influence capital allocation, risk assessment, and strategic planning. Building them on untested assumptions is a form of analytical negligence. The durbin watson test is a 30-second check that can prevent months of decisions built on flawed math.

Ready to put this knowledge to work? Try Warren, your AI financial advisor, get personalized, conflict-free guidance at heywarren.com