Combination Forecast - Which models to pick?

Question

Combination Forecasting can be produced by simply averaging different forecasts or employing more complex techniques (see Makridakis, 1989; De Gooijer and Hyndman, 2006; Goodwin, 2009; Pesaran and Pick, 2011). The major principle of forecast averaging is simply by computing the average of two forecasts for the forecasting period.

I plan to do a combination forecast for real estate cycle prediction. As you can see above I read a lot of literature about it, but cannot yet decide, which models I choose to implement in R?

Any recommendation, on which factors a combination forecast can be decided?

I appreciate your replies!

Answer 1

Folklore among forecasters is (to my knowledge) that it is usually empirically better to combine "very different" models. So instead of combining all possible exponential smoothing forecasts, it may be better to combine an exponential smoothing forecast, an ARIMA forecast (noting that some ARIMA and smoothing algorithms are more-or-less equivalent), a neural network forecast, a simple benchmark and a judgmental forecast. Or, in your case where smoothing is likely not very good for real estate cycle prediction, use different econometric models, a neural network and a number of different analysts' judgmental forecasts. However, I can't back this folklore up with references.

It seems like a simple unweighted combination, possibly after trimming or winsorizing point forecasts, may already be pretty hard to beat by more complex weighting schemes (Kolassa, 2011 - note shameless self-promotion - and Claeskens et al., 2016 on the "forecast combination puzzle"). If you include judgmental forecasts, it may even be hard to determine weights for weighting schemes, since judgmental forecasts don't come with information criteria, in-sample fits, degrees of freedom and so forth.

Answer 2

There are two questions, I would answer them seperately:

• I plan to do a combination forecast for real estate cycle prediction. As you can see above I read a lot of literature about it, but cannot yet decide, which models I choose to implement in R?
  • Forecasting cycle is one of the most complex tasks for any forecasters. There is no general model that one could recommend for your specific problem. As Armstrong notes in his article NOT to forecast cycles as they are stochastic in nature and lack predictability or unless you know (e.g., based on contractual relationships or on physical or biological laws) that cycles will occur and have good knowledge about timing. See also this article by same author (p14). You could try forecasting by analogue. simple illustrative example follows. See below the picture from Japanese real estate bubble analogues with US home price bubble (Source: Signal and Noise by Nate Silver). Coincidentally this is similar to what you are asking about real estate cycle. The above 2 article also have excellent information on combining forecasts.

• Any recommendation, on which factors a combination forecast can be decided?
  • In general combining forecasts works very well. I would not stop recommending the article by Armstrong on combining forecast that provides general and excellent guidance for forecasting practitioners. This would supplement Diebolds excellent chapter on combining forecast techniques that @RichardHardy has referenced. No researchers that I know of, has summarized extensive research on combining forecasts like Armstrong. Below is a screenshot summary of the article. I'm including this because sometimes the link to the article can be broken, and it might not be useful for future readers.

Hope this helps