Checking Residual Normality
What you'll learn: How to verify that your time series model's residuals follow a normal distribution—a key assumption for valid inference.
Why Normal Residuals Matter
After fitting an ARIMA model, you've already tested that residuals are independent (using the Ljung-Box test) and inspected their autocorrelation structure. Now comes another important check: are the residuals normally distributed?
If residuals approximate a normal distribution, it validates your model's assumptions and ensures that confidence intervals and hypothesis tests are trustworthy. Non-normal residuals might indicate outliers, model misspecification, or the need for transformations.
Three Ways to Check Normality
1. Histogram
Plot a histogram of your residuals and overlay a normal curve. If residuals are roughly bell-shaped and symmetric, that's a good sign. Heavy skewness or multiple peaks suggest problems.
2. Q-Q Plot (Quantile-Quantile Plot)
This is your visual gold standard. A Q-Q plot compares your residuals' quantiles against theoretical normal quantiles. If points fall roughly along a straight diagonal line, residuals are approximately normal. Deviations at the tails indicate non-normality.
Analogy: Imagine lining up students by height (your residuals) and comparing them to a "perfect" height lineup from a normal distribution. If they match well, the line is straight.
3. Statistical Tests
Formal tests (like the Shapiro-Wilk or Jarque-Bera test) give you a p-value. A high p-value (typically > 0.05) suggests residuals don't significantly deviate from normality. However, these tests can be overly sensitive with large datasets.
In Practice
Use all three methods together. The Q-Q plot is particularly informative because it shows where normality breaks down (e.g., heavy tails). Minor deviations are often acceptable; ARIMA models are fairly robust to slight non-normality.
Key Takeaway: Checking residual normality through histograms, Q-Q plots, and statistical tests ensures your time series model's assumptions hold, making forecasts and inference more reliable.