Time series analysis to forecast revenue

[requejo14]

Hi,

I work in OpFin for a search agency. Our clients work with 3 different platforms (Google, Yahoo and Microsoft) and 2 of them (Yahoo! and Microsoft) are merging next year in Europe (Both platforms already merged in the US back in 2010)

I am working on a project to forecast European revenue once migration occurs (Q3 2012) and I have thought about using an autoregression model using US figures as my raw data to build an autoregression model that hopefully captures that change in dynamics that the new marketplace brings. Once I have my US regression model I plan to apply it to EU data. (For the US I have 2 years worth of data for both platforms prior to the merge and 1 year in the new marketplace)

I have a number of concerns:

- I don't expect the time series to be co-variance stationary
- Seasonality is a factor to take into account when forecasting quaterly revenue.

How would I best transform the data to cope with the above issues?

Thanks in advance

[vinux]

what variable you are analyzing? what is the frequency?

[requejo14]

Revenue (US dollars) and traffic (clicks), I have daily figures but provided the model holds I am planning to agregate figures to monthly or quaterly frequencies.

Cheers

[vinux]

Hi,

I have a number of concerns:

- I don't expect the time series to be co-variance stationary
- Seasonality is a factor to take into account when forecasting quaterly revenue.

How would I best transform the data to cope with the above issues?
Thanks in advance

This would be an interesting data. Probably you just need to do a log transformation. I assume Revenue is positive. The transformation is decided based on the stationarity condition. See the time series plot ( visual check) then do stationarity test on original series and log series. Visual part gives an idea of what transformation to be applied. But that is mostly an art( experience) and trial and error method.

To get an covariance part, see the cross correlations.

Ideally minimum 3 years of data required for estimating seasonal part. But still can estimate seasonality using 2 years.

[noetsi]

You could also apply ARIMA but that takes even more experience (with the method and the data). It too is an art form.