I understand ARMA is a linear combination of lagged data points and lagged errors, but I am unclear on its implementation once parameters have been identified. Now suppose I have an ARMA model and some data. Where do I get these error terms, especially for the first term.