Hi guys! While I was doing some questions of the Time-Series Analysis part, I came up with a question which I can’t fully justify de answer. The question says:
Which of the following AR models is most appropiate for a time series with annual seasonality using quarterly observations?
a) b1xt-1 + b2xt-12 + et
b) b0 + b1xt-1 + b2xt-4 + et
b) b0 + b1xt-4 + b2xt-12 + et
The correct answer is B. Can you explain why is it please?
I think the answer is C, rather than B. Answer B incorporates monthly and quarterly trends, assuming each step in time is 1 month. Answer C is correct then because it incorporates quarterly and yearly trends if each step in time is 1 month (meaning t-12 is the prior year, 12 months ago).
Where did this question and answer come from? I can say the question isn’t worded as cleanly as I would expect, either, so I may be misinterpreting some of their unclear word choice.
The question wording is clear enough, it just tricked your mind.
We must pay attention to two things:
How the data is built: daily, weekly, monthly, quarterly, annual, etc
What is the suspected seasonality. The question ask for annual seasonality. There are tests for error autocorrelation that can show up to 36-lag analysis, so if we see an error pike in the chart, we can find the lag level of the suspected seasonality.
An AR model for a time series is commonly built in the simplest form :
X(t) = a + bX(t-1) + e
If this form does not work well, then we can make adjustments. The most common adjustment is seasonality. In this case, an annual seasonality in a quarterly data would be built as:
X(t) = a + bX(t-1) + cX(t-4) + e
because each quarter has 3 months and 4 x 3 = 12 (annual).