The purpose of the project is for students to write a report on forecasting a variable of their choice. Hence there are two distinct and important aspects of this:
The objective is creating a forecast model of a variable of choice, rather than creating a model that forecasts well. Intrinsic uncertainty in economic variables ensures that forecasting well is not necessarily possible, and even a good forecast model may forecast badly on a small number of observations.
The first objective, thus, is the thorough selection of a good forecasting model. This involves making sensible choices at all stages of the process, and being transparent about those choices and their likely impact. This involves being thorough in the selection of a forecast model, detailing why particular decisions were made, and implementing all of the methods covered in class, where necessary, in the pursuit of a good forecasting model.
For example, rather than relying on auto.arima(), a better project will detail the steps outlined in the textbook and lecture slides in selecting an ARIMA model, using the tsdisplay() function to make sense of the patterns in the data. A better project will spend a lot of time exploring what explanatory variables might be useful, given the constraint that a forecast model must be created (hence not all explanatory variables are feasible). A better project will attempt to make as balanced as possible a comparison between forecast models on a range of criteria, and explore why particular types of models appear to work better than others.
Not all projects will be amenable to a range of models like those covered in class being implemented; in such a case, it must be explained in detail why this is the case, and more detail devoted to the construction of this one single forecast model. For example, if simulation methods are employed, these should be explained in detail.
Similarly, not all datasets will involve the same level of work to manipulate into something from which a forecasting model can be created. It is expected that in such a situation, a range of forecast outputs will be considered, rather than one particular type of forecast, such as a one-step ahead foercast. For example, longer-term forecasts might be produced, and evaluated using a training sample, as might forecast methods that forecast explanatory variables using different methods.
The second objective is to write an accurate and thorough account of the forecasting process. Such an account will be faithful to the process, recording the important decisions made along the way. This is an important distinction, and one that is difficult to make — it can be hard to know what is important, and what is not. Two tips on this:
.R file is helpful here). Then go through the list, and think about which ones you imagine had the most influence on the outcomes of your forecasting exercise, and include them. The rest, I would suggest devote an appendix to, which you might refer to in your footnotes.As well as being faithful to what you did, the report ought to also be:
On (1), ensure the text is broken up by objects such as graphics and tables. My suggestion would be avoid having too much code in the text; I would suggest having in the text only the code relevant for the graphics and modelling; the detailed manipulation necessary to get the dataset into place should be included in the appendices. On this, it is important to note that manipulating data pre-R in Excel, whilst comfortable as you know Excel well, is less amenable to recording every decision in the way a .R file could. My strong suggestion is to do as much data manipulation and preparation in .R as you possibly can; anything you can do in Excel you can certainly do in R, and lots more besides.
On (2), talk about your data, describe it, show you know every detail of it so that the reader (examiner) is convinced you know the data best, and are best placed to model it. Then, of course, reflect that discussion of the data in your subsequent modelling choices. As in, don’t talk about possible structural breaks and then omit to carry out any structural break analysis.
Anything less than a 1st class project fails on one of the above accounts.
Start writing immediately, using R Markdown (.Rmd) to embed code and graphics in (do not include graphics you created elsewhere, or in R, as graphics files — create them in the .Rmd file). It doens’t matter if you write too much, simply write as you go along, this will help you to collect your thoughts.
Try to Google things; copy and paste in the error message, and the command you’re using. Try to make sense of the output, and try alternative ways to do what you’re trying to do (in R!). If all else fails, email me. I won’t be able to reply before April 8 necessarily, but I will try.