Time Series and Forecasting
For this project, you are to build a data set of at least two, but no more than four, variables that you believe might be related over time. It’s probably best to use macroeconomic data (national or state level). Good sources of US macro data include the FRED database at the St. Louis Federal Reserve Bank, and the data at the Bureau of Labor Statistics (BLS.gov).
Your data should be monthly, with a minimum of 15 years (180 observations). I would prefer at least 20 years, but I know of some data that only go back to 2000, so I’ll accept 15. If your data are released daily or weekly, you’ll have to use XL to get the daily (or weekly) averages for the month.
You are to provide two forecasts of the final year of your data: a VAR of all of your variables together, and ARIMA(p,d,q) forecasts of each variable separately. You may use either the entire data set to choose your models, or you may truncate it not to include the final year. Once you’ve decided on a model, run the model with the truncated data to generate your forecasts. In the VAR case, you’ll need to explain why you chose the number of lags you used, and if there was predictive causality between your variables (and in what direction: e.g., X had some for Y, but Y didn’t for X). In the ARIMA models, you’ll need to document why you chose the final structures, and show any unit root tests you might have run.
Once you have your forecasts, compare them to the actual. Which method resulted in forecasts with the smallest MSE? Briefly explain why you believe that method worked better.
When it is time to turn your project in, please submit the RATS program as well as your write up. It’s best to print any graphs generated by RATS as you go, for once you close the program, the graphs disappear.
Links to the data to be used: