Difficiculties with predicting GDP

[Maria Petersen]

Hi, I am writing my bachelor's thesis right now, and I am having difficulties with the predict command.. I have read all the manuals and "help" tab as well but I don't seem to get it right. My dataset is panel data for a time period of 2000-2019. I have appended the time period to 2025, but I don't seem to get how I can predict the gdp for future years. I only seem to get values for the years that I have data on all the explanatory variables, which is an issue since I am studying Kenya, and there is a lot of missing data... so I have two questions:

1. How do i predict the years of missing values in the explanatory variables? (For example, I have data for 2000-2007 but no data for 2008-2010, and then data exists until 2018 again. I need to fill in the blanks)

2. How do i predict future years of gdp? since I don't have any values for the explanatory variables during these future years as well..?

I hope that the explanation of my problem is not too unclear, english is not my first language..

[Maria Petersen]

Almost forgot to add, I have tried using the 'forecast' command instead, but I got the error message that said: "command may not be used with panel data".. so that didn't work..

[Mike Lacy]

While there might be ways to make -predict- work in your situation, I'd think that -forecast- is more relevant. I haven't used -forecast-, but examining -help forecast- I find the phrase "forecast works with both time-series and panel datasets," so I'd suspect the difficulty comes from something about how you tried to use -forecast- rather than the command itself. Because you didn't show the actual command you used, you have made it difficult (perhaps impossible) for someone to help you. Please read the FAQ for new StataList members, and re-post showing a sample of your data (-help dataex-) and your exact code. I don't think know enough to help you here, but with the right information from you, someone else could.

[Oscar Ozfidan]

Maria, the industry practice (what the consultants or forecasters do) is usually they fit the model like you have and they purchase forecasts of explanatory variables from a company or another source and they append those to the data so that the predict would produce those results in a case like yours. If that is the route you choose, using arima or exponential smoothing (these require only the history of the independent variable of interest, which you would have to do one at a time).

There is also VAR option (that is vector auto-regression). In your case, I think the easiest way would be to use VAR. VAR would produce dynamic forecasts of all endogenous variables based on their past correlations. That said having a gap in your time series would likely be a problem but for a bachelor thesis I would (if I were your professor) accept if you filled those missing data for those years with their expected values or imputed them. So, you might want to check with your advisor if that would fly. Or what approach he would accept filling data for those years. If he would not accept anything other than actual data, then you cant use data prior to 2011 and that would not I am afraid leave enough data for the VAR to work.