Course contentsShow
← Data Science
Lesson 782 of 2,14520. Time Series ModelsFree lesson

Residual Diagnostics for ARIMA

Check residual autocorrelation, normality, and homoscedasticity to validate model assumptions.

Residual Diagnostics for ARIMA

You'll learn how to verify your ARIMA model is properly specified by examining the residuals it leaves behind.

What Are Residual Diagnostics?

After fitting an ARIMA model, you need to check whether it captured all the patterns in your data. The residuals (the differences between actual and fitted values) should behave like pure random noise if your model is good. If they don't, your model missed something important.

Think of it like cleaning a window: if you've done a thorough job, only random water droplets remain. If you still see streaks or patterns, you need to keep cleaning (or respecify your model).

Three Key Diagnostic Checks

1. Residual Autocorrelation

Plot the ACF of residuals. Ideally, all lags should fall within the significance bounds—meaning no leftover correlation structure. If significant spikes appear, your ARIMA orders (p, d, q) may be wrong.

2. Normality

ARIMA assumes residuals are normally distributed. Use a histogram or Q-Q plot to visually check, or apply formal tests like Shapiro-Wilk. Non-normal residuals suggest outliers or model misspecification.

3. Homoscedasticity

Residuals should have constant variance over time. Plot residuals vs. time or fitted values. If you see a "funnel" shape (variance increasing/decreasing), you may need transformations or a different model structure.

Why This Matters

Good residual diagnostics confirm your model has extracted all useful information. Poor diagnostics mean biased forecasts and unreliable confidence intervals. These checks validate the assumptions underlying ARIMA—just like you learned for GLMs, but now applied to time series.

Key Takeaway: Residuals from a well-specified ARIMA model should resemble white noise: no autocorrelation, normally distributed, and constant variance across time.