# Time series analysis on small dataset

I am new to time series and therefore trying to feel my way around. I am required to do some analysis and possibly come up with a model. However, I have at most 2 years worth of quarterly data (at most 8 datapoints). Can I build any meaningful model based on the data? What possible model apart from linear regression with time as regressor can be built?

Thanks

• Not really. Eight data points wouldn’t even be enough for simple ordinary least squares regression, or even a t-test. I’d like to ask, what is it that you are trying to accomplish? What is your research question?
– Jon
Aug 2, 2018 at 5:20
• A graph would be useful (no need of the units). If you follow Tuker's advice, data visualization should become an automatic reflex. Aug 2, 2018 at 11:05

There is no valid answer. If you have a stationary process with 0 variance, then the forecast horizon has no limit.

More realistically, you have the follow rule of thumb (which is totally from experience with absolutely no theoretical base): the horizon forecast may be half the historical base. You have 8 data points, so you can forecast 4 points. You have 2 years, so you can forecast 1 year.

The key point is : check this after your forecast is done.

• Is there any mathematical proof for ´the horizon forecast may be half the historical base´or it is just based on experience? Jan 4, 2020 at 16:19
• A. Debcker, T. Modis, Determination of the uncertainties in S-curve logistic fits, Technological Forecasting and Social Change Volume 46, Issue 2, June 1994, Pages 153-173 Apr 4, 2020 at 8:11

As far as the time evolution model is concerned, you have:

• A linear growth/decline. You increase/decrease of a constant value (in units, dollars,...)

• A exponential growth/decline. You increase/decrease of a constant rate (in %)

• A logistic growth/decline. You are filling a limited niche, and the increase in % is proportional to what remain to be filled.

• Higher level of model (Bass-diffusion models,...), approximations (polynomials,...) or adaptive methods (moving average, exponential smoothing, ARIMA,...)

You can easily test exponential growth by fitting the log of your data on a linear regression, but other models require more sophisticated techniques.

• Why do you not combine your two answers into one, e.g. using two heading, if you are making two different points? Aug 2, 2018 at 11:03

I would start playing around with the prophet library for python, you can still make predictions with 8 datapoints, the accuracy is going to depend on a lot of things especially the variance in the points.

• Suggesting someone try a statistical package does not address the issue of data quality/quantity which is OPs concern.
– Jon
Aug 2, 2018 at 5:21