Ardl forecasting

[Ardeshir Saedi]

• Hi, Dear All,
  
  I estimated an ARDL model, in which "Budget Deficit/GDP" is dependent variable and two most important independent variables are GDP GAP & G/GDP.
  
  All observations are Yearly , and time series contain more than 50 years annual budget deficit and ....
  
  All tests done and model is quite acceptable.
  
  Now having the outlook of GDP GAP and G/GDP for next five years, (as independetn variable) I'm trying to forecast the "Budget Deficit/GDP" for each next five years.
  
  I used commands as below
  1- ARDL ......... , aic
  2- estimate store model
  3- forecast creat model
  4- forecast estimate model
  5- forecast exogenous .....
  6- forecast solve, prefix (_est) end (2027)
  
  Outputs generated by software, and up to now every thing is o.k.
  
  But I'm not sure about accuracy,
  cause, the generated/forecasted Budget Deficit ratio turned very soon ( just after one or two next year) to the long term trend of itself,
  
  sine ARLD characteristic enjoys a long term principals to show how variables have cointegrated behaviour in long term. I'm wondering should I use another command (or trick) to get right forecast for a short term 5 years outlook? In this regard
  
  question-1: should I add "ec" after "aic" option at the end of ARDL command and then should use the coefficients reported in SR section? To reach precision forecast for next five years? Which I think 5 years must obey a short term dynamics comparing to structural coefficient that my model estimates based on 50 years observation.
  
  Question-2: as I add ec to ARDL command and repeat command from 1 to 6 (above mentioned) new problem arises,
  "forecast solve" process terminates immidiately after first year saying "missing data for forecasting 2024"!! To solve it I added new identity to model as
  budget deficit/gdp=d.(budget deficit/gdp) + L.(budget deficit/gdp)
  Now forecasting continues but output remains constant!!
  
  would be very grateful if you advise me how to cope with such issue.
  
  And please accept my excuse for this long and primitive question writing here.
  
  Ardeshi

[Ardeshir Saedi]

GDP GAP & G/GDP = GDP GAP & G/GDP

[Sebastian Kripfganz]

Depending on the estimated coefficients, it might well be the case that the dependent variable returns to its long-run path relatively quickly. I would not add the ec option because that would give you forecasts for the changes, not the levels.

I am not very experienced with the forecast command suite. Thus, I am afraid I am unable to help much further.

[Ardeshir Saedi]

Dear Sebastian,
I would like to thank you very much,

as you know when we add "ec" to options, Stata will report speed of adjustment and LR and SR separately.

If I want to calculate forecast "by hand", is it right to use SR coefficients in excel sheet and then by having explanatory variables?

Could it be more reliable for five years (short run) forecast?

LR coefficients is built based on the 50 years observation (time series) so I suppose should not be considered for five year forecast.

would be appreciated to have your advise

my best
Ardeshir

[Sebastian Kripfganz]

The adjustment back into the equilibrium is still considered part of the short-run dynamics. Therefore, I would not leave out the long-run equilibrium relationship and the corresponding speed-of-adjustment coefficient, unless the conclusion from the bounds test is that there is no long-run relationship.