I know it might look naive but I have a very basic question. I have a three years of historic data which has weekly and annual seasonality. January first as my first data is on Wednesday so my time series starts from Wednesday not Monday. Does it affect on my forecasting model? I mean if the model considers the right seasonality for days of a week if the first point of my data is a day in the middle of the week not the first day of the week?(I saw that for monthly data series, the data always starts from January which is the begining of the season, should daily data sets follow the same logic and start from Monday?)
3
-
$\begingroup$ I don't see why the first day being a Wednesday should be any problem; it's not always the case that monthly data starts in January. $\endgroup$– Glen_bCommented Sep 17, 2014 at 15:44
-
$\begingroup$ I confess that Im new to forecasting and my knowledge is limited. I saw some datasets which were started in January but you are definitely right and that might not be the case always. $\endgroup$– user12Commented Sep 17, 2014 at 15:46
-
$\begingroup$ If you're still concerned, run one particular model with the data filtered to start on a monday and verify that there is little change. I agree with others that it shouldn't make a difference. $\endgroup$– Chris UmphlettCommented Sep 5, 2018 at 15:50
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
It normally shouldn't affect your forecasting model whether your weeks go from monday to sunday (most of Europe), from sunday to saturday (US?) or from wednesday to tuesday (your case). You might make sure that your model doesn't implicitely add something like a weekend effect.
-
$\begingroup$ Thanks for your information. I have daily seasonality which mean that weekend have lower values and weekdays have higher values. Hence weekend are definitely affecting the model. What should I do in this case? I simply wrote this line of code to show time series: x <- msts(orders, seasonal.periods=c(7,365.25)), this shows that I included multiple seasonality. when reviewing fitted values, weekend values are not close to the actual values so I thought maybe daily seasonality is not working properly in my model $\endgroup$– user12Commented Sep 17, 2014 at 13:17
-
$\begingroup$ msts doesn't consider weekends special per se, so that shouldn't be a problem. Are yousure that every seventh value corresponds to a weekend (i.e. no gaps in the data)? Can you see at least some weekly trend in the forecast? $\endgroup$– dobiwanCommented Sep 17, 2014 at 13:59
-
$\begingroup$ I dont have any missing data and I predicted 30 days out of sample. It definitely shows a simillar trend as the historic data but the values for weekend predictions are almost 50% lower than the actual value. weekdays are pretty fin and the prediction error for them is less than 10%. BTW my best fitted model is as follows: TBATS(1, {4,3}, 1, {<7,2>, <365.25,6>}) $\endgroup$– user12Commented Sep 17, 2014 at 14:06