0

Currently I am studying time series Moving Average model MA(q) $$X_t -\mu= \epsilon_t + \theta_1\epsilon_{t-1} + \theta_2 \epsilon_{t-2} + ... + \theta_q \epsilon_q$$ where $\theta_1,...,\theta_q$ are parameters and $\epsilon_t$ is a white noise with mean $0$ and constant variance.

I have been puzzling on how to compute $\epsilon_t.$ Are they just residuals?

This post illustrated how to compute $\epsilon_t$ in $MA(1)$ model. But I do not see how the same technique can be carried forward to $MA(q)$ model.

Idonknow
  • 291
  • 3
  • 12

1 Answers1

1

" Are they just residuals? " . Yes !

Read my response to Moving-average model error terms

IrishStat
  • 27,906
  • 5
  • 29
  • 55
  • In this case, what is a procedure of fitting data with MA(q) model? Which variable should we find first? As residual is the difference between true value and predicted value by forecasting model, I suppose residual is calculated the last. Since we can obtain sample acf, by theoretical formula of acf of MA(q) model, we can find variable $\theta_1,...,\theta_q.$ Am I right? – Idonknow Jul 22 '19 at 13:38
  • the coefficients are found by optimization i.e. trial & error through iteration. – IrishStat Jul 22 '19 at 13:50
  • I see. But the trial and error stage is done first? Then we proceed to estimate other parameters? – Idonknow Jul 22 '19 at 13:51
  • trial and error is done for all parameters simultaneously – IrishStat Jul 22 '19 at 13:52
  • But I thought there is a method to choose them efficiently... – Idonknow Jul 22 '19 at 14:18
  • no they are estimated based upon the model and the data – IrishStat Jul 22 '19 at 15:17