Eric, thank you so much for making this video & providing so many additional links! I felt super lost regarding a class project of forecasting for which I have 6 days to complete without any prior knowledge of how to write forecasting code, but you give me hope! :)
Thank you so much Eric for your essential presentation about how to forecast GDP. In the meantime, I kindly ask that could you share your do-file if it is possible?
@@econ_eric hi Eric, you mentioned squared deviation from mean in the video and say there is much deviation from mean corresponding to the crisis, I am not sure if the squared deviation here referring to the variance
@@yuqingwu1011 I'm not sure where I said squared deviation in the video, but the GDP growth associated with a crisis will be a large deviation from the mean of the GDP growth series.
Hi Eric amazing work here..! I like to ask regarding the lag aspect. As i run through your code, i realized that the Q-value is not less than 0.05, and also the regression output for p-value as well. Just like to ask if it would be still meaningful to regress gdp growth to l.gdp growth? As im guessing the whole idea pins upon the lagged aspect. Thanks!
Yes! It is still meaningful to regress gdp growth on l.gdp growth if it produces accurate forecasts. I would argue that the relationship is not spurious - a good year of growth is likely to spur another good year. Consider the alternatives and determine which model produces more accurate forecasts. I was thinking about making another video about lag specification or checking the accuracy of forecasts. There are several ways to check the accuracy of your forecasts - one easy way is to compare the actual data to the forecast. The Stata manual for the forecast command explains this well beginning on page 6: www.stata.com/manuals13/tsforecast.pdf If the coefficient on the lag of GDP growth is not significantly different from zero with 0.05 level of significance, we can try another model with only the regression constant or perhaps another variable or two. Then, plot and compare the forecasts between models, compare r-squared values, etc. In any case, if a model accurately forecasts, then it is hard to argue that it is not meaningful because a p-value for a coefficient is 0.054 rather than
@@econ_eric thanks for the comprehensive comment! A lot to digest and yes a video regarding lag specification and forecasting accuracy sound like good content! Thanks for the reply :)
Hi Eric, another question , If the variable is not stationary, but after first difference the dfuller shows it is stationary, and then the best lag is lag 2, what should the reg command be? (The original variable is x, after first difference it is x1, and dfuller shows the optimal lag for x1 is lag2) , really confusing about this…
If you dfuller shows stationarity (reject Ho), then however you set up the dfuller test should be your base model. Then you adjust things. Check out p.2 of the tsdfuller pdf. If you're trying to determine which model is the best model to forecast with, then I would generate a new dependent variable series, delete some of the most recent observations, and forecast with each model. Some models may forecast a recovery from recession better and others will forecast the growth before a downturn better. Which is more important to you? There are also the indicators of model fit to consider. Also, I'm not sure if you mean the optimal lag # or the optimal lag. Check out the tsdfuller pdf for details on the output of dfuller and what you're actually testing.
Also, you can test whether a model is stationary by predicting residuals (postestimation in Stata). Predict the residuals and then use dfuller to determine if the residuals are stationary. Do this for each variation of your model. Also, If you find that the first difference of your variable is stationary, then try including a one-period lag of the variable on the right side of the equation (as an independent variable). Predict the residuals and determine if they are stationary. This can be a viable alternative to using your first difference model from dfuller as your main model.
Same issue here. I saved the coefficient estimate from regression model. But the error still states "Missing values were encountered while attempting to solve the model at time 2023. Variable NUTHIV evaluates to missing." What might I be doing wrong? Thank you beforehand!
@@econ_eric Yes I know there is no panel data in your video. I ask if you have the command to do this graph that figures in your video at the second 0.04 and also at the time 8.48 minute (forcasting graph) but with panel data case. I need to get this graph for my paper using panel data approach.
Samia, what is your panel variable? If it identifies countries, then do you want real GDP growth plotted for each country in your data set? The variable f_realgdpgrowth includes a forecast of GDP growth, also. Therefore, many options are possible for an estimation and plot.
Hi Eric, thank you for your video. I go through your do file and it looks like "forecast solve" only works when the model is "reg realgdpgrowth l.realgdpgrowth". If we want to add more independent variables there (for example reg realgdpgrowth l.realgdpgrowth worldgdpgrowth), then how can we have the out of sample forecast? I tried your do-file but it does not work. Can you please help me ?
Hello Duong, thank you for your comment. I think your solution is to store the estimates from the new model with worldgdpgrowth and then use the forecast command. For instance, include worldgdpgrowth in the model on line 79, store the estimates, and then create a new forecast model on line 96 (call it something different than ericmodel). Please let me know if that works.
