0

I have quarterly data from 2017 to 2020 (16 data points) to forecast. I tried to use SARIMA but it is giving me weird numbers (High and negative fitted values). Also, I tried exponential smoothening which is working better than SARIMA. But 2020 data points is affecting the forecast (Checked against 1st quarter real data for 2021).

I need few suggestions here -

  1. Which method would be best for less data points?

  2. Should I transform (disaggregate) quarterly to monthly/weekly? is it recommended?

  3. Since SARIMA is not working fine this data and exponential smoothening doesn't allow to add exogenous regressor (by research I came to know), I couldn't able to add any covid related variable to my model. any suggestion how I can add exogenous regressor.

  4. Considering data has seasonality exponential smoothening seems to be working okay but because of 2020 covid effect it is affecting future forecast.

  5. Can I remove 2020 data points from my model (train dataset) to achieve good accuracy?

suggestion would be highly appreciated.

Feel free to comment if any more information required to help.

Arpit Soni
  • 1
  • 1
  • 16 points is a very small dataset, you should probably use much simpler methods https://stats.stackexchange.com/questions/135061/best-method-for-short-time-series – Tim Jun 23 '21 at 11:23

1 Answers1

0

If you have just 16 data points I would not suggest removing any more data points. You will lose p+s (lag order + seasonalities) degrees of freedom just for fitting your model. Furthermore, you mention there is an impact from covid in 2020 in your data, this can be interpreted as either noise in your data, or worse, a structural break in your series. Both are detrimental for your forecast (accuracy). Just to be clear, noise and a small sample will impact your forecast accuracy, not the forecast value itself.

As far as your first suggestion goes, disaggregating your data to lower frequency, will increase your small sample size. This might seem helpful, because it helps your estimation. But it really is basically waving a magic wand and create data that will exactly glue your initial time series together. This is not recommended.

Lastly, if you have bad data you can do only so much. There is no model or method of estimation that will heal your bad data. My recommendation is to estimate a simple model (maybe just an AR(1)?) and don't put too much stock in your prediction.

confused student
  • 451
  • 1
  • 2
  • 8