0

I am trying to forecast the regional GDP growth of our region in the next five years, I only have 20 observations in my data which is yearly, what forecasting model is appropriate?

I tried ARIMA in r and the forecast.ets function in excel. The ARIMA results to (0,0,0) with non-zero mean which gave me the same forecast points so I tried to suppress the mean. Below is the table showing the results I got:

ARIMA Exponential smoothing
2021 -1.10 0.44
2022 -0.64 5.24
2023 -0.37 -0.12
2024 -0.21 -0.14
2025 -0.12 0.41

Here is the plot for the original time series and ACF+PACF:

enter image description here

And here is the plot for the exponential smoothing:

enter image description here

Anisah
  • 1
  • 2
  • Could you add a plot of the original time series and the ACF + PACF? – Adrià Luz Sep 21 '21 at 09:50
  • 1
    "ARIMA(0,0,0) with non-zero mean" means that `auto.arima()` sees no structure whatsoever in your data. Its best guess is that your observations are IID. Then the best forecast is just the historical average. Which may indeed be the best you can do with just 20 years of data. – Stephan Kolassa Sep 21 '21 at 10:16
  • @AdriàLuz I added the plot – Anisah Sep 21 '21 at 10:32
  • @StephanKolassa I have read that using ARIMA to short time series would only give me very bad forecast, I am still trying to find what forecasting method is suitable but I'm still honestly lost at the moment. – Anisah Sep 21 '21 at 10:37
  • Unfortunately, 20 data points is just extremely little data for fitting any model, especially since the main drivers of GDP are external factors (financial crisis of 2008, COVID in 2020). No non-causal model can account for these correctly. Your best bet is likely the historical average, or some reasonable judgmental forecasts. Or a scenario analysis. – Stephan Kolassa Sep 21 '21 at 10:40
  • @StephanKolassa I see, what about naive forecasting? I have read in some blogs this method can be also used in short time series. – Anisah Sep 21 '21 at 11:16
  • Yes, that is also a possibility. However, if you just forecast the last observation out, you are completely at the mercy of that last observation. In 2020, we had COVID. Do you really want to use the 2020 data point to forecast out the next five years? – Stephan Kolassa Sep 21 '21 at 11:31
  • @StephanKolassa thanks a lot, I think I will just go with the scenario analysis – Anisah Sep 22 '21 at 01:25

0 Answers0