2
$\begingroup$

I tried to find similar questions on this topic but couldn't find anything that helped me with my problem, so I will try to explain it on my own.

I'm trying to do a regression analysis on revenue for the past 6 years (not taking 2020 into account, due to corona-related differences). The data I have on our revenue is on a daily basis, but considering the independent variables I use (e.g. GDP) and their availability I aggregated them up to quarters. Considering the fact that i only have 6*4=24 observations, the regression won't be too accurate I guess, but that can't be changed, so I have to ignore that.

My problem is now, that I got quite confused/lost in trying to find out, how to do my analysis.

The revenue and my other variables are time series, so they mostly aren't independent on their own. To counter that problem I tried to find a solution, but only found some regarding the time aspect. With that I mean seasonality and things like that. The revenue of course has some seasonality, in the 3rd quarter there is always a peak for the year.

So here is my question: How can I do a regression analysis on revenue, regarding the time factor and also different other variables like GDP, "manager's 6 month expectation"-index,... (using SPSS)?

Is it correctly done, if I let SPSS analyze autocorrelation and then create lagged variables for significant and reasonable lags, which I include as a new independant variable? (Just as an example based on what I found so far)

Some things that might help to know:

  • My dependant variable is the revenue of my Company, the independant ones are factors like GDP, produced cars, Leadtime of certain products and so on
  • Overall I have around 10 variables, which is also due to 3 being dummies for quarters
  • I used Excel first for some Tests but now started with SPSS
  • On my first attempts I got "wrong" results, like our revenue having a negative (but significant) correlation to the GDP, which doesn't make sense as it usually follows the GDP but the general trend of the revenue is negative and the trend of the GDP is positive > therefore I probably need to find other/better variables
  • On my first attempts I also tested for seasonality, which led to the result, that there is no significant change for quarters, even though it is easily identifiable if you look at it on a Chart
  • My experience in analysis is kind of restricted, as I only know some theory I learned in my statistics class in university and never used it in a practical case before
Improve this question
$\endgroup$
9
  • $\begingroup$ This is very unclear. What are you measuring ? You say "our revenue", so is this a single company and you want to do a regression of the company revenue vs GDP and some other variables ? $\endgroup$
    – Robert Long
    Commented Oct 28, 2020 at 20:24
  • $\begingroup$ Yes, sorry. I want to do a regression on a single companys (my company) revenue with GDP and some other variables. $\endgroup$
    – I'm a questionmark
    Commented Oct 29, 2020 at 8:33
  • $\begingroup$ OK so why would it be a mixed effects model ? What would be the random effects ? $\endgroup$
    – Robert Long
    Commented Oct 29, 2020 at 10:08
  • $\begingroup$ That's one of my Problems, because i don't know an answer to that. I tried finding a way to run my Analysis and one of the things I found, was a recommendation, to use a mixed-model. But by the way you're asking this question, i guess that this wouldn't help me at all. Still, random effects might be factors, that influence the behaviour of our customers, that I can't predict. For example, if they had an accident and some of the parts we deliver broke, so they order new ones. $\endgroup$
    – I'm a questionmark
    Commented Oct 29, 2020 at 10:46
  • $\begingroup$ No that doesn't really make sense. You would use random effects (random intercepts) when you have repeated measures or some other form of clustering of observations. So if you were analysing the assoiation of some variable on revenue across multiple companies then you would fit random intercepts for company, $\endgroup$
    – Robert Long
    Commented Oct 29, 2020 at 11:41

2 Answers 2

Reset to default
0
$\begingroup$

Yes, this would generally be handled as a single time series. SPSS has a procedure named TSMODEL that includes an Expert Modeler function to try to choose the best model. TSMODEL is accessed in the menus via Analyze>Forecasting>Create Models or Create Traditional Models, depending on the release (older releases might not have the "Traditional" included). First you'll want to clarify the time structure using Data>Define Dates and specifying quarterly data. Then you don't need the seasonal indicators, as the program already knows the data are seasonal.

Improve this answer
$\endgroup$
0
$\begingroup$

On my first attempts I got "wrong" results, like our revenue having a negative (but significant) correlation to the GDP, which doesn't make sense as it usually follows the GDP but the general trend of the revenue is negative and the trend of the GDP is positive > therefore I probably need to find other/better variables

This is not wrong. With time series you can find all kinds of correlations. What is wrong is to draw strong conclusions with these type of data. See also: Why do these time series appear to be dependent? and Udny Yule's "Why do we Sometimes get Nonsense-Correlations between Time-Series?--A Study in Sampling and the Nature of Time-Series".

Is it correctly done, if I let SPSS analyze autocorrelation and then create lagged variables for significant and reasonable lags, which I include as a new independant variable?

You could do something like that. You could do a Granger causality test which analyses whether changes in one series can be predicted by changes in the other series, and this will take autocorrelation into consideration. With only 6x24 points you may probably not have much significance, and also the interpretation remains difficult (the relationship found can be spurious).

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.