Practical process
- Inspect plot. If mean-variance relationship, use variance stabilising transformation if necessary: square root (for count data, or mild relationship) or log (for strong relationship)
- If clearly non-stationary difference, otherwise inspect ACF for very slow decay. Difference until this is gone. Check at seasonal intervals as well
- Plot ACF and PACF for differenced series. Identify as AR, MA, or ARMA. If AR or MA get p or q from cutoff. Check seasonal lags as well. In ARMA first try p=1 and q=1.
- Estimate parameters.
- If any non-significant reduce model and estimate again.
- Plot diagnostics (if data is seasonal make sure to extend for multiple seasonal lags). If problems try adding extra terms to model.