Error while fitting data in auto.arima - R

Question

I am running auto.arima for forecasting time series data and getting the following error:

1: The time series frequency has been rounded to support seasonal differencing. 
2: In value[[3L]](cond) :
  The chosen test encountered an error, so no seasonal differencing is

selected. Check the time series data.

This is what I am executing:

fit <- auto.arima(data,seasonal = TRUE, approximation = FALSE)

I have weekly time series data.

This is how dput(data) looks as follows:

structure(c(12911647L, 12618317L, 12827388L, 12967840L, 13264925L, 
13557838L, 13701131L, 13812463L, 13971928L, 13837658L, 13550635L, 
13022371L, 13507596L, 13456736L, 12992393L, 12831883L, 13262301L, 
12831691L, 12808893L, 12726330L, 11893457L, 12434051L, 12363464L, 
12077055L, 12107221L, 11986124L, 11997087L, 12264971L, 12164412L, 
12438279L, 12733842L, 12543251L, 12627134L, 12480153L, 12276238L, 
12443655L, 12497753L, 12279060L, 12549138L, 12308591L, 12416680L, 
12516725L, 12326545L, 12772578L, 12524848L, 13429830L, 14188044L, 
16611840L, 16476565L, 15659941L, 10785585L, 12150894L, 13436366L, 
12985213L, 13097555L, 13204872L, 13786040L, 13760281L, 13295389L, 
14734578L, 15043941L, 14821169L, 14361765L, 14300180L, 14357964L, 
14271892L, 13248168L, 13813784L, 14092489L, 14100024L, 13378374L, 
13225650L, 12582444L, 13267163L, 13026181L, 12747286L, 12707074L, 
12534595L, 12546094L, 13030406L, 12950360L, 12814398L, 13405187L, 
13277755L, 13142375L, 12742153L, 12610817L, 12267747L, 12570075L, 
12704157L, 12835948L, 12851893L, 12978880L, 13104906L, 12754018L, 
13213958L, 13584642L, 13963433L, 14471672L, 16312595L, 16630000L, 
16443882L, 11555299L, 12018373L, 13031876L, 13013945L, 13164137L, 
13313246L, 13652605L, 13803606L, 13308310L, 14466211L, 15092736L, 
15346015L, 14467260L, 14767785L, 13914271L, 14185070L, 13851028L, 
13605858L, 13597999L, 13876994L, 13026270L, 13113250L, 12288727L, 
12925846L, 13525010L, 12594472L, 12654512L, 12888260L), .Tsp = c(2016.00819672131, 
2018.48047598209, 52.1785714285714), class = "ts")

This is how I am reading data from the csv

read_data <- read.csv(file="data.csv", header=TRUE)
data_ts <- ts(read_data, freq=365.25/7,  start=decimal_date(ymd("2016-1-4")))
data <- data_ts[, 2:2]

This is the data in the csv:

Year    si_act
1/4/16  12911647
1/11/16 12618317
1/18/16 12827388
1/25/16 12967840
2/1/16  13264925
2/8/16  13557838
2/15/16 13701131
2/22/16 13812463
2/29/16 13971928
3/7/16  13837658
3/14/16 13550635
3/21/16 13022371
3/28/16 13507596
4/4/16  13456736
4/11/16 12992393
4/18/16 12831883
4/25/16 13262301
5/2/16  12831691
5/9/16  12808893
5/16/16 12726330
5/23/16 11893457
5/30/16 12434051
6/6/16  12363464
6/13/16 12077055
6/20/16 12107221
6/27/16 11986124
7/4/16  11997087
7/11/16 12264971
7/18/16 12164412
7/25/16 12438279
8/1/16  12733842
8/8/16  12543251
8/15/16 12627134
8/22/16 12480153
8/29/16 12276238
9/5/16  12443655
9/12/16 12497753
9/19/16 12279060
9/26/16 12549138
10/3/16 12308591
10/10/16    12416680
10/17/16    12516725
10/24/16    12326545
10/31/16    12772578
11/7/16 12524848
11/14/16    13429830
11/21/16    14188044
11/28/16    16611840
12/5/16 16476565
12/12/16    15659941
12/19/16    10785585
12/26/16    12150894
1/2/17  13436366
1/9/17  12985213
1/16/17 13097555
1/23/17 13204872
1/30/17 13786040
2/6/17  13760281
2/13/17 13295389
2/20/17 14734578
2/27/17 15043941
3/6/17  14821169
3/13/17 14361765
3/20/17 14300180
3/27/17 14357964
4/3/17  14271892
4/10/17 13248168
4/17/17 13813784
4/24/17 14092489
5/1/17  14100024
5/8/17  13378374
5/15/17 13225650
5/22/17 12582444
5/29/17 13267163
6/5/17  13026181
6/12/17 12747286
6/19/17 12707074
6/26/17 12534595
7/3/17  12546094
7/10/17 13030406
7/17/17 12950360
7/24/17 12814398
7/31/17 13405187
8/7/17  13277755
8/14/17 13142375
8/21/17 12742153
8/28/17 12610817
9/4/17  12267747
9/11/17 12570075
9/18/17 12704157
9/25/17 12835948
10/2/17 12851893
10/9/17 12978880
10/16/17    13104906
10/23/17    12754018
10/30/17    13213958
11/6/17 13584642
11/13/17    13963433
11/20/17    14471672
11/27/17    16312595
12/4/17 16630000
12/11/17    16443882
12/18/17    11555299
12/25/17    12018373
1/1/18  13031876
1/8/18  13013945
1/15/18 13164137
1/22/18 13313246
1/29/18 13652605
2/5/18  13803606
2/12/18 13308310
2/19/18 14466211
2/26/18 15092736
3/5/18  15346015
3/12/18 14467260
3/19/18 14767785
3/26/18 13914271
4/2/18  14185070
4/9/18  13851028
4/16/18 13605858
4/23/18 13597999
4/30/18 13876994
5/7/18  13026270
5/14/18 13113250
5/21/18 12288727
5/28/18 12925846
6/4/18  13525010
6/11/18 12594472
6/18/18 12654512
6/25/18 12888260

Edit1 :

I was able to read the data without any errors before, initially, I had 160 records & the model does not throw any error but, then for 80-20 test I removed the last 30 records and this error cropped up. Now also, if I run with all the data I don't get any error but is I run it with first 130 as 80% I get this error.

Answer 1

This warning (not an error) is informing you that the seasonal unit root test (used to select the number of seasonal differences, D) has errored.

Admittedly, the message is not very informative for why this has happened. In your case, an STL decomposition cannot be performed because your data contains less than two seasonal windows after the first seasonal difference. This is necessary to use the nsdiffs(y, test = "seas"), or auto.arima(y, seasonal.test = "seas") which are both defaults.

Note that the message informs you that D=1 has been selected for this data. The test for D=2 failed due to insufficient available data (generally one seasonal difference is sufficient, however you should check this visually or with other unit root tests).

I've now improved this message to now also include the error message for why the test has failed: https://github.com/robjhyndman/forecast/commit/eebea5ee93cd8b125d5220c54721895b57396157

library(forecast)
fit <- auto.arima(data,seasonal = TRUE, approximation = FALSE)
#> Warning: The time series frequency has been rounded to support seasonal
#> differencing.
#> Warning: The chosen seasonal unit root test encountered an error when testing for the second difference.
#> From stl(): series is not periodic or has less than two periods
#> 1 seasonal differences will be used. Consider using a different unit root test.
fit
#> Series: data 
#> ARIMA(0,1,2)(0,1,0)[52] 
#> 
#> Coefficients:
#>           ma1      ma2
#>       -0.5560  -0.2496
#> s.e.   0.1124   0.1110
#> 
#> sigma^2 estimated as 1.199e+11:  log likelihood=-1090.71
#> AIC=2187.42   AICc=2187.74   BIC=2194.45

^{Created on 2019-04-24 by the reprex package (v0.2.1)}