1

I have a time series with temperature measures every 5-minutes over ca. 5-7 days. I'm looking to set the correlation structure for my model as I have considerable temporal autocorrelation. I've decided that moving averages would be the best form, but I am unsure what to specify within the correlation = corARMA(q=?) part of the model. Here is the following output for ACF(m1):

       lag          ACF
1    0  1.000000000
2    1  0.906757430
3    2  0.782992821
4    3  0.648405513
5    4  0.506600300
6    5  0.369248402
7    6  0.247234208
8    7  0.139716028
9    8  0.059351579
10   9 -0.009968973
11  10 -0.055269347
12  11 -0.086383590
13  12 -0.108512009
14  13 -0.114441343
15  14 -0.104985321
16  15 -0.089398656
17  16 -0.070320370
18  17 -0.051427604
19  18 -0.028491302
20  19  0.005331508
21  20  0.044325557
22  21  0.083718759
23  22  0.121348020
24  23  0.143549745
25  24  0.151540265
26  25  0.146369313

It appears that there is highly significant autocorrelation in the first ca. 7 lags. See also the attached images: 1[Residuals] & 2[Model]

Would this mean I set the correlation = corARMA(q=7)?

Crow
  • 11
  • 4
  • can't tell you if your suggested code is correct. But the R package "sandwich" has a function to estimate robust standard errors in the presence of heteroscedasticity and autocorrelation to compute coefficients. I don't know enough about statistics to compare the two approaches. If you have the knowledge I'd be interested to know your take on it :-) – Simone Jun 08 '18 at 13:51
  • In case you haven't seen this yet https://rdrr.io/cran/nlme/man/corARMA.html – Simone Jun 08 '18 at 15:43

0 Answers0