Written and maintained by CASRAI Editorial Board
Published
On this page
The Jarque-Bera test is an asymptotic goodness-of-fit statistical test designed to determine whether sample data or regression residuals match the skewness and kurtosis of a normal distribution. Developed by econometricians Carlos Jarque and Anil Bera, the test evaluates the joint null hypothesis that sample skewness equals zero and sample kurtosis equals three. Under this null hypothesis of normality, the test statistic follows a chi-square distribution with two degrees of freedom, making it an essential diagnostic tool for validating distributional assumptions in econometrics, financial modeling, and time series regression analysis.
Definition of the Jarque-Bera Test
The Jarque-Bera test evaluates whether an empirical dataset or regression residual series departs significantly from Gaussian normality based on two central standardized shape parameters: skewness and kurtosis. In statistical modeling and econometrics, parametric procedures such as Ordinary Least Squares (OLS) regression rely on normality assumptions to establish exact finite-sample distributions for standard errors, t-statistics, and F-tests. When evaluating the normality of a distribution, empirical researchers require rigorous numerical criteria rather than informal graphical inspections alone.
Unlike distance-based tests such as the Kolmogorov-Smirnov test that evaluate the maximum vertical discrepancy between empirical and theoretical cumulative distribution functions, the Jarque-Bera procedure focuses directly on the third and fourth standardized central moments. A Gaussian distribution possesses perfect symmetry, resulting in a theoretical skewness of zero. Furthermore, it exhibits a mesokurtic profile with a theoretical kurtosis of three. The Jarque-Bera test formalizes these properties into a joint composite hypothesis:
- Null Hypothesis (H0): The data or residuals are normally distributed, with skewness S = 0 and kurtosis K = 3 (excess kurtosis K – 3 = 0).
- Alternative Hypothesis (H1): The data or residuals depart from normality, where skewness S ≠ 0, kurtosis K ≠ 3, or both conditions hold.
Because the test evaluates skewness and kurtosis simultaneously, it produces a single omnibus test statistic. If the observed sample exhibits noticeable skewness, excessive kurtosis, or both, the test statistic escalates. If the value exceeds the critical threshold of a chi-square distribution, researchers reject the null hypothesis and conclude that the underlying data generating process departs from normality.
Jarque-Bera Test Formula and Mathematical Foundations
The Jarque-Bera test statistic, conventionally denoted as JB, combines standardized sample skewness and sample kurtosis into a single quadratic score. The mathematical formula is expressed as:
JB = (n / 6) × [S2 + ((K – 3)2 / 4)]
Each mathematical symbol in the formula represents a specific parameter of the sample:
- n: The sample size, representing the total number of valid observations or regression residuals used in the estimation.
- S: The sample skewness, which quantifies the degree of asymmetry of the distribution around its sample mean. It corresponds to the standardized third central moment: S = m3 / (m2)3/2, where mr = (1/n) ∑ (Yi – Ȳ)r. In symmetric distributions, S equals zero. Positive values reflect right-skewed data with an elongated right tail, whereas negative values reflect left-skewed data. For foundational mechanics on asymmetry, see our guide to skewness in data distributions.
- K: The sample kurtosis, representing the standardized fourth central moment: K = m4 / (m2)2. Kurtosis measures the propensity of the distribution to generate tail values and outliers relative to its center. For a Gaussian distribution, K equals three.
- (K – 3): The excess kurtosis. Subtracting three centers the kurtosis metric around zero for a normal distribution. A distribution with positive excess kurtosis (K > 3) is leptokurtic, possessing heavier tails and a sharper central peak. A distribution with negative excess kurtosis (K < 3) is platykurtic, featuring thinner tails.
- 6 and 24 (via 6 × 4): Normalizing constants derived from the asymptotic variances of sample skewness and kurtosis under the null hypothesis of normality.
The theoretical derivation of the test traces to the Lagrange Multiplier (LM) score test principle introduced by Carlos Jarque and Anil Bera in 1980 and formalized in 1987. The authors framed the test within Pearson’s general family of frequency curves, which nests normal, Student’s t, beta, gamma, and Pareto distributions. Under the Pearson system, setting specific shape parameters to zero yields a normal distribution.
Jarque and Bera demonstrated that under the null hypothesis of normality, the standardized sample moments converge asymptotically to independent normal variables:
- √n × S → N(0, 6) as n → ∞
- √n × (K – 3) → N(0, 24) as n → ∞
Because sample skewness and kurtosis are asymptotically uncorrelated under normality, standardizing each term produces independent standard normal z-scores: zS = √n × S / √6 and zK = √n × (K – 3) / √24. Squaring these two independent standard normal variables yields two independent chi-square variables with one degree of freedom each:
(zS)2 = (n / 6) × S2 ~ χ2(1)
(zK)2 = (n / 24) × (K – 3)2 = (n / 6) × [((K – 3)2 / 4)] ~ χ2(1)
Because the sum of two independent chi-square variables with one degree of freedom produces a chi-square variable with two degrees of freedom, the composite statistic converges in distribution: JB → χ2(2). Contextualizing this distribution alongside other common test metrics is discussed in our overview of types of test statistics.
Worked Example With Step-by-Step Arithmetic
To illustrate the practical calculation of the Jarque-Bera test, consider an empirical econometric study testing whether the residuals of an asset-pricing model meet the assumption of normality. The model produces a series of regression residuals across a sample of 120 monthly observations.
The sample calculations from the raw residual series provide the following summary parameters:
- Sample size (n): 120
- Sample skewness (S): 0.35
- Sample kurtosis (K): 4.20
- Significance level (α): 0.05
The test statistic and statistical significance are computed through the following arithmetic steps:
Step 1: Compute the Skewness Component
Calculate the square of the sample skewness S:
S2 = (0.35)2 = 0.1225
The positive skewness indicates a modest rightward tail in the residual distribution.
Step 2: Compute Excess Kurtosis and the Kurtosis Component
Determine excess kurtosis by subtracting 3 from sample kurtosis K, then square the result and divide by 4:
Excess Kurtosis = K – 3 = 4.20 – 3.00 = 1.20
(K – 3)2 = (1.20)2 = 1.4400
((K – 3)2 / 4) = 1.4400 / 4 = 0.3600
The positive excess kurtosis confirms a leptokurtic distribution with heavier tails than expected under normality.
Step 3: Sum the Moment Components
Combine the squared skewness and the scaled kurtosis component:
Sum of Components = S2 + ((K – 3)2 / 4) = 0.1225 + 0.3600 = 0.4825
Step 4: Scale by the Sample Size Factor
Multiply the bracketed sum by the sample size scaling factor (n / 6):
(n / 6) = 120 / 6 = 20.0000
JB = 20.0000 × 0.4825 = 9.6500
The computed Jarque-Bera test statistic is exactly JB = 9.65.
Step 5: Determine the Critical Value and P-Value
Under the null hypothesis of normality, JB is evaluated against a chi-square distribution with two degrees of freedom (χ2 with df = 2). For a chi-square distribution with df = 2, the survival function has the closed-form expression p = exp(-JB / 2):
- Critical Value Method: At α = 0.05 with df = 2, the critical value from standard tables is χ20.05, 2 = 5.991. Because the computed statistic JB = 9.65 exceeds 5.991, the null hypothesis is rejected.
- P-Value Method: Evaluating the survival function yields p = exp(-9.65 / 2) = exp(-4.825) ≈ 0.00803. Because p = 0.00803 is strictly less than α = 0.05 (and less than α = 0.01), the departure from normality is statistically significant at the 1% significance level.
Substantive Conclusion: The researcher rejects the null hypothesis of residual normality at both the 5% and 1% levels. The residuals exhibit significant non-normality driven primarily by leptokurtosis (fat tails), indicating that hypothesis testing must account for potential size distortions.
Interpretation of Test Results and Decision Guidance
Interpreting Jarque-Bera test results involves balancing mathematical hypothesis decisions with practical modeling consequences. Because the JB statistic is strictly non-negative, any positive departure from zero reflects divergence from normality. Values close to zero support the plausibility of Gaussian assumptions, whereas large values lead to rejection.
The decision guidance table below summarizes standard statistical thresholds and practical modeling responses:
| Jarque-Bera Statistic (JB) | P-Value (χ2, df=2) | Null Hypothesis (H0) Decision | Distributional Assessment | Recommended Practical Action |
|---|---|---|---|---|
| 0.00 to 4.60 | p ≥ 0.10 | Fail to reject H0 | No statistically significant departure from normality detected. | Proceed with standard parametric inference, t-tests, and F-tests. |
| 4.61 to 5.99 | 0.05 ≤ p < 0.10 | Fail to reject at 5%; reject at 10% | Marginal evidence of non-normality; minor skewness or kurtosis. | Inspect Q-Q plots; evaluate sample size robustness and outlier influence. |
| 6.00 to 9.21 | 0.01 ≤ p < 0.05 | Reject H0 at 5% level | Statistically significant non-normality; noticeable asymmetry or kurtosis. | Use robust standard errors, bootstrap confidence intervals, or transformations. |
| > 9.21 | p < 0.01 | Reject H0 at 1% level | Strong, highly significant non-normality; severe tail weight or asymmetry. | Avoid relying on finite-sample normality; employ robust regression or GMM. |
In regression analysis, researchers must distinguish between coefficient unbiasedness and hypothesis testing validity. Under the Gauss-Markov theorem, OLS estimators remain the Best Linear Unbiased Estimators (BLUE) even when error terms are non-normal, provided errors are homoscedastic and uncorrelated. Non-normality does not introduce bias into estimated regression slopes.
However, non-normality directly affects hypothesis testing in small samples. When disturbances deviate from normality, t-statistics and F-statistics do not strictly follow Student’s t or Snedecor’s F distributions in finite samples. In large samples, the Central Limit Theorem ensures that regression coefficients converge toward asymptotic normality regardless of error shape. Nonetheless, when sample sizes are modest or extreme outliers generate heavy tails, standard t-tests can exhibit size distortions, producing inflated false-positive rates.
Assumptions, Limitations, and Common Mistakes
Applying the Jarque-Bera test effectively requires awareness of its theoretical limits, computational subtleties, and common analytical mistakes.
1. Asymptotic Sample Size Requirement
The foremost constraint of the Jarque-Bera test is that its reference distribution is strictly asymptotic. The test statistic converges to a chi-square distribution with two degrees of freedom only as sample size n approaches infinity. In finite samples, the sampling distribution of sample kurtosis converges slowly to its asymptotic variance of 24/n.
Statistical computing documentation, including SciPy guidelines, highlights that the Jarque-Bera test works best for large sample sizes, ideally exceeding 2,000 observations. In small samples (e.g., n < 50 or n < 100), the asymptotic test can exhibit noticeable size distortions and reduced power. For smaller samples, specialized procedures such as the Shapiro-Wilk test in R offer substantially higher statistical power and accurate size control.
2. The Excel Kurtosis Definition Trap
A widespread computational error occurs when implementing the Jarque-Bera formula in Microsoft Excel. In standard statistical notation, K represents raw kurtosis (where a normal distribution equals 3), and the formula explicitly subtracts 3: (K – 3). However, Excel’s built-in KURT() function calculates excess kurtosis, which has already subtracted 3 from the raw fourth moment. If an analyst enters ((KURT(range) - 3)^2) / 4, they subtract 3 twice, corrupting the kurtosis term and producing an invalid test statistic. When using Excel’s KURT(), the term must simply be squared directly: (KURT(range)^2) / 4.
3. Conflating Failure to Reject With Proof of Normality
Obtaining a p-value greater than 0.05 indicates that the sample lacks sufficient evidence to reject the normal distribution at that significance level. It does not establish that the population is strictly normal. In small samples, the test lacks power, often failing to reject normality even when data originate from moderately skewed or bimodal distributions. Researchers should always supplement statistical tests with graphical diagnostics, including normal Q-Q plots and kernel density estimates.
4. Sensitivity to Single Outliers
Because skewness relies on the third power of deviations and kurtosis relies on the fourth power, the Jarque-Bera statistic is exceptionally sensitive to extreme outliers. A single erroneous data point or measurement error can inflate sample kurtosis, causing the test to reject normality even if the remainder of the dataset conforms to a Gaussian distribution. Researchers should examine residual plots and leverage statistics to determine whether non-normality reflects genuine distributional shape or isolated data contamination.
5. Independent and Identically Distributed (I.I.D.) Assumption
The Jarque-Bera test assumes that observations are independent and identically distributed. When applied to regression residuals, the presence of serial correlation or heteroscedasticity violates this assumption, distorting the variance of sample moments and invalidating the χ2(2) reference distribution. When serial correlation is suspected in time series models, autocorrelation diagnostics must be evaluated concurrently.
Comparison With Alternative Normality Tests
The Jarque-Bera test is one of several goodness-of-fit procedures used in empirical research. Choosing the appropriate test depends on sample size, whether parameter values are known or estimated, and whether the primary research focus is tail behavior, asymmetry, or omnibus distributional shape.
The data table below compares the statistical properties, recommended sample sizes, and primary sensitivities of common goodness-of-fit normality tests:
For general statistical applications with sample sizes below 2,000 observations, the Shapiro-Wilk test remains the gold standard omnibus test due to its superior power across a vast range of non-normal alternatives. For testing empirical cumulative distribution functions where mean and variance are estimated from data, the Lilliefors correction to the Kolmogorov-Smirnov test or the Anderson-Darling test is preferred. The Jarque-Bera test remains the predominant standard in econometrics and finance because it directly tests the moment conditions that underpin regression theory and financial risk metrics.
How to Report Jarque-Bera Test Results in APA Style
When presenting quantitative findings in academic papers adhering to American Psychological Association (APA 7th edition) style, statistical test results must follow clear reporting conventions. Statistical abbreviations and symbols such as JB, p, and N are italicized, degrees of freedom are declared, and exact p-values are reported to two or three decimal places (without a leading zero before the decimal point).
Because the Jarque-Bera statistic asymptotically follows a chi-square distribution with two degrees of freedom, researchers may report it either using the test-specific symbol JB or standard chi-square notation χ2 with degrees of freedom in parentheses:
Example 1: Reporting Non-Normality (Null Rejected)
“A Jarque-Bera test on the Ordinary Least Squares regression residuals revealed a statistically significant departure from normality, JB = 9.65, p = .008, driven primarily by positive excess kurtosis (K = 4.20). Consequently, heteroscedasticity-and-autocorrelation-consistent (HAC) robust standard errors were utilized for all subsequent parameter inference.”
Alternatively, using chi-square notation:
“Evaluation of model residuals indicated significant deviation from Gaussian distribution assumptions, χ2(2, N = 120) = 9.65, p = .008.”
Example 2: Reporting Plausible Normality (Null Not Rejected)
“The regression residuals conformed satisfactorily to distributional normality assumptions, as confirmed by a non-significant Jarque-Bera test statistic, JB = 1.42, p = .492 (skewness S = 0.08, kurtosis K = 3.24). Standard parametric hypothesis testing procedures were therefore retained.”
How to Calculate the Jarque-Bera Test in Excel, R, and Python
Modern statistical workflows allow researchers to execute the Jarque-Bera test easily across multiple platforms. Below are tested, production-ready code snippets and formulas for Excel, R, and Python.
Calculating in Microsoft Excel
Although Excel lacks a single dedicated JARQUE.BERA function, the statistic is easily calculated using native worksheet formulas. Assuming raw observations or regression residuals reside in cells A2:A121:
- Sample Size (n):
=COUNT(A2:A121) - Sample Skewness (S):
=SKEW(A2:A121) - Excess Kurtosis (K – 3):
=KURT(A2:A121)(Reminder: Excel’s KURT returns excess kurtosis directly) - Jarque-Bera Statistic (JB):
=(COUNT(A2:A121)/6) * (SKEW(A2:A121)^2 + (KURT(A2:A121)^2)/4) - P-Value:
=CHISQ.DIST.RT(D5, 2)(where cellD5contains the computed JB statistic)
Calculating in R
In the R statistical environment, the test is provided in the widely used tseries package via jarque.bera.test(), and can also be implemented using base R:
# Method 1: Using the tseries package
install.packages("tseries")
library(tseries)
# Fit a linear regression model
model <- lm(yield ~ rating + maturity, data = bond_data)
# Run the Jarque-Bera test on model residuals
jb_output <- jarque.bera.test(residuals(model))
print(jb_output)
# Method 2: Standalone base R function without external dependencies
jarque_bera_calc <- function(x) {
n <- length(x)
m_diff <- x - mean(x)
m2 <- mean(m_diff^2)
m3 <- mean(m_diff^3)
m4 <- mean(m_diff^4)
s <- m3 / (m2^(3/2))
k <- m4 / (m2^2)
jb <- (n / 6) * (s^2 + ((k - 3)^2) / 4)
p_val <- pchisq(jb, df = 2, lower.tail = FALSE)
return(list(statistic = jb, p_value = p_val, skewness = s, kurtosis = k))
}
# Execute calculation
res <- jarque_bera_calc(residuals(model))
cat(sprintf("JB Statistic: %.4f | p-value: %.5fn", res$statistic, res$p_value))
Calculating in Python
In Python, the Jarque-Bera test is available in both scipy.stats and statsmodels:
import numpy as np
from scipy import stats
import statsmodels.api as sm
from statsmodels.stats.stattools import jarque_bera
# Example residuals array
np.random.seed(42)
residuals = np.random.normal(loc=0.0, scale=1.0, size=2500)
# Approach 1: scipy.stats.jarque_bera
res = stats.jarque_bera(residuals)
print(f"SciPy -> JB: {res.statistic:.4f}, p-value: {res.pvalue:.5f}")
# Approach 2: statsmodels.stats.stattools.jarque_bera
jb_stat, p_val, skew, kurt = jarque_bera(residuals)
print(f"Statsmodels -> JB: {jb_stat:.4f}, p-value: {p_val:.5f}, Skew: {skew:.4f}, Kurt: {kurt:.4f}")
# In Statsmodels OLS regression, the JB test is included automatically:
# X = sm.add_constant(predictors)
# model = sm.OLS(y, X).fit()
# print(model.summary()) # JB and Prob(JB) are displayed under Residual Diagnostics
Frequently asked questions
What is the null hypothesis of the Jarque-Bera test?
The null hypothesis (H0) states that the sample data or regression residuals are normally distributed, having a population skewness of zero and a population kurtosis of three. The alternative hypothesis (H1) states that the data deviate from a normal distribution due to non-zero skewness, non-three kurtosis, or both.
How large must the sample size be for the Jarque-Bera test to be valid?
Because the Jarque-Bera test relies on asymptotic convergence to the chi-square distribution with two degrees of freedom, it requires relatively large samples to ensure reliable test size and power. Statistical computing literature, including SciPy documentation, recommends sample sizes of at least several hundred observations, and ideally exceeding 2,000 observations. For smaller samples (n < 100), the test can exhibit significant size distortions, and tests like Shapiro-Wilk are strongly recommended.
Why does the Jarque-Bera test have 2 degrees of freedom?
The test evaluates two independent parameter restrictions simultaneously: whether skewness equals zero (one restriction) and whether kurtosis equals three (a second restriction). Under the null hypothesis of normality, standardized sample skewness and kurtosis converge asymptotically to two independent standard normal variables. Summing their squared values produces a chi-square distribution with two degrees of freedom (df = 2).
What should you do if the Jarque-Bera test rejects normality in regression residuals?
If residual normality is rejected, first determine whether non-normality is driven by isolated outliers, model misspecification (omitted non-linearities), or genuine heavy tails. Depending on the diagnosis, appropriate corrective actions include applying non-linear transformations (such as logarithmic or Box-Cox transformations) to dependent variables, utilizing heteroscedasticity-and-autocorrelation-consistent (HAC) standard errors, or employing non-parametric bootstrapping for confidence intervals.
How does the Jarque-Bera test differ from the Shapiro-Wilk test?
The Jarque-Bera test is a moment-based test that specifically evaluates whether third and fourth standardized moments match Gaussian expectations, relying on asymptotic theory. In contrast, the Shapiro-Wilk test is an order-statistic test that evaluates the correlation between ordered sample observations and theoretical normal quantiles. Shapiro-Wilk has substantially higher statistical power in small-to-moderate samples (n ≤ 2,000), whereas Jarque-Bera is predominantly favored in econometrics and large-sample time series modeling.
Can the Jarque-Bera test statistic be negative?
No, the Jarque-Bera test statistic cannot be negative. The formula calculates the sum of two squared terms multiplied by a positive sample size factor (n / 6). Because both S2 and ((K – 3)2 / 4) are squared real numbers, they are non-negative. Consequently, the minimum possible value of the JB statistic is zero, which occurs only if sample skewness is exactly zero and sample kurtosis is exactly three.
Sources
- Jarque, C. M., & Bera, A. K. (1987). A Test for Normality of Observations and Regression Residuals. International Statistical Review, 55(2), 163-172. https://doi.org/10.2307/1403192 (Checked 2026-11-14)
- Jarque, C. M., & Bera, A. K. (1980). Efficient tests for normality, homoscedasticity and serial independence of regression residuals. Economics Letters, 6(3), 255-259. https://doi.org/10.1016/0165-1765(80)90024-5 (Checked 2026-11-14)
- National Institute of Standards and Technology (NIST). Section 1.3.5.11: Measures of Skewness and Kurtosis. NIST/SEMATECH e-Handbook of Statistical Methods. https://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm (Checked 2026-11-14)
- SciPy Community. scipy.stats.jarque_bera Reference Documentation. SciPy Reference Guide. https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.jarque_bera.html (Checked 2026-11-14)
- Statsmodels Developers. statsmodels.stats.stattools.jarque_bera Documentation. Statsmodels Manual. https://www.statsmodels.org/stable/generated/statsmodels.stats.stattools.jarque_bera.html (Checked 2026-11-14)
- Trapletti, A., & Hornik, K. (2024). Package ‘tseries’: Time Series Analysis and Computational Finance. Comprehensive R Archive Network (CRAN). https://cran.r-project.org/web/packages/tseries/tseries.pdf (Checked 2026-11-14)
- Urzúa, C. M. (1996). On the simplicity and speed of robust tests for normality. Economics Letters, 51(1), 1-8. https://doi.org/10.1016/0165-1765(95)00778-5 (Checked 2026-11-14)
Data: Comparison of Statistical Normality Tests in Empirical Research
Comparative specifications of five primary goodness-of-fit normality tests used in statistical modeling, detailing their asymptotic distributions, recommended sample size domains, primary moment sensitivities, and standard computational package implementations compiled from statistical literature and software documentation.
| Test Name | Test Statistic Symbol | Reference Distribution | Degrees of Freedom | Minimum Recommended Sample Size (observations) | Primary Sensitivity | Key Software Implementation |
|---|---|---|---|---|---|---|
| Jarque-Bera Test | JB | Chi-square | 2 | 2000 | Skewness and kurtosis deviations | scipy.stats.jarque_bera / tseries::jarque.bera.test |
| Shapiro-Wilk Test | W | W distribution (tabulated) | None | 3 | Broad omnibus sensitivity across entire distribution shape | scipy.stats.shapiro / stats::shapiro.test |
| Anderson-Darling Test | A^2 | Empirical distribution function | None | 5 | Tail behavior and extreme values | scipy.stats.anderson / nortest::ad.test |
| D'Agostino-Pearson K^2 Test | K^2 | Chi-square | 2 | 20 | Transformed standardized skewness and kurtosis | scipy.stats.normaltest / moments::agostino.test |
| Lilliefors (Kolmogorov-Smirnov) Test | D | Lilliefors distribution table | None | 4 | Maximum vertical distance from empirical CDF with estimated parameters | statsmodels.stats.diagnostic.lilliefors / nortest::lillie.test |
Sources
- Jarque & Bera (1987) A Test for Normality of Observations and Regression Residuals (accessed 2026-11-14)
- NIST/SEMATECH e-Handbook of Statistical Methods (accessed 2026-11-14)
- SciPy Reference Guide: scipy.stats.jarque_bera (accessed 2026-11-14)
- CRAN R Package tseries Reference Manual (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







