ARIMA Model Estimation
What you'll learn: How ARIMA parameters are fitted to your time series data using maximum likelihood estimation.
From Candidate to Fitted Model
You've identified candidate ARIMA orders (p, d, q) using ACF and PACF plots following the Box-Jenkins methodology. Now comes the crucial step: estimation—finding the exact coefficient values that make your model best fit the data.
Maximum Likelihood Estimation (MLE)
ARIMA models use maximum likelihood estimation to find optimal parameters. Think of MLE as asking: "Which coefficient values make the observed data most probable?"
Here's the intuition: Your time series is like a partially completed puzzle. The ARIMA structure gives you the puzzle's shape (AR terms, MA terms, differencing), but MLE finds the specific piece sizes (coefficient values) that make all the observed pieces fit together most naturally.
MLE works by:
- Proposing coefficient values
- Calculating how likely your actual data would be under those values
- Adjusting coefficients to maximize that likelihood
- Repeating until no improvement is possible
What Gets Estimated
For an ARIMA(p,d,q) model, MLE estimates:
- φ coefficients for AR terms (p values)
- θ coefficients for MA terms (q values)
- σ² for the error variance
- Sometimes a constant/intercept term
The differencing order (d) is applied before estimation—it's not estimated, you chose it.
Why It Matters
Without proper estimation, even a correctly identified ARIMA structure won't forecast well. MLE ensures your model captures the true data-generating process as closely as possible given your chosen structure.
Key Takeaway: ARIMA estimation uses maximum likelihood to find coefficient values that make your observed time series most probable, transforming a model structure into a fully specified forecasting tool.