Thanks for your great video! But one question regarding your explanation: I don't think the only potential term equaling zero is Exp(Error(t-1)^2), instead it should be one Error(t-i) within k
6:54 in the Auto-correlation term: Why you aren't taking in consideration in the second term E(Xt)*E(Xt-k) ?! Shouldn't it be auto-correlation is diffrent from 0 if the first term is diffrent from μ^2 ?
That's true but that's not what the statement's saying (note that it's NOT equal to 0). The statement says that the only way to get (at least) two terms in common is for k less than or equal to q.
@@utpalpodder-pk6vq I think you are right. The condition in the video is wrong. He was trying to show there is no overlap, and if there is no overlap there will be no common term present, and if there are no common term present, it should be = 0 and not != 0. So if it is = 0 (no overlap), then the last term E[X(t-q)] should not overlap the first term E[X(t-k)]. Hence, t-q should be >= t-k instead.
If the MA model uses the errors from the previous periods to forecast, why are we not using the PACF (which is the correlation of residual over the actual values) to determine the appropriate q for the MA model?
Great video but from around 6:30 onwards your words do not match the equations you write. You are saying in words that the inequality between t-q and t-k leads to the overall expected value being zero when it actually leads to the overall expected value being NON-zero as in the equals sign with a cross through it on the left. Took me a while to figure out what you were saying.
To further clarify, if the very last term in X_t with time interval t-q is SMALLER than the t-k value of the very first term in X_t-k, then there is identical error variable overlaps and hence an overall non-zero expected value. On paper what you write makes complete sense but your words say the opposite.
Hi great videos..really helping a lot..i am just starting data science and economics course...can u please help by making videos on basiscs like wold decomposition, invertibility, impulse response,Linear fiters and forecasting
hi there - can you give any guidance on the method used to fit MA(q) processes - i.e. find the phi parameters. I can't find much information about this
Oh, it's using the entire covariance equation: E[xy] - E[x]E[y]. The mu^2 gets cancelled out.
3 года назад
@@scottpease9827 Hi, I was wondering the same .. do I understand it correctly that: E[xy] = mu^2 (if the errors are not overlapping) and E[x]E[y] = mu^2, then: E[xy]-E[x]E[y]=mu^2-mu^2=0?
Why expectation (Xt,Xt-k) is not zero for most terms as except first term all other terms have error terms? As you said expectation of error will always be equal to zero.
Thx for the video ritvikmath, i have one question ( i might missed it in the video). Video tells us why ma(3) dont use et-4 term bcs its autocorr is 0 thus it wont add anything to our model. Just like AR vs PACF logic. is that correct? So this is like an explanation video for "why we dont use all the lags and cut the formula after k=q?" I hope i made myself clear. Have a great day.
A basic question comes to mind I.e if the expected value of error term is zero, then why at all include the error terms in the time series prediction . In case the EV of Error is not zero, then can the EV value be straight away added to time series prediction without doing all the correlation calculations
While plotting graph of ACF or PACF against lag ...you talked about error band ....so what is the range of error band we should take.... I mean what are the parameters of error band we should consider..... please reply fast