Im currently doing research using some macroeconomic data.
I just want to know what are the pros and cons of using seasonally adjusted/non-seasonally adjusted data in terms of forecasting.
$\begingroup$
$\endgroup$
1
-
2$\begingroup$ substract the seasonality, do the forecast and then add the seasonality :) (this is only relevant for data with a frequency of less than a year, otherwise you do not need to do SA). $\endgroup$– luchonachoOct 12, 2017 at 6:21
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
There is a theoretical objection to using seasonally adjusted data, which I saw in an essay by Kalman (developer of the Kalman filter); sorry, I cannot find the reference. The argument is that seasonal adjustment is a filter, that has its own dynamics. You are then embedding these filter dynamics within the system that you are modelling. This would be a bad idea in engineering systems, but I am not entirely convinced it matters for economic systems. (My academic background is in engineering control systems.) However, if the variables you are forecasting exhibit roughly the same seasonal pattern, separately seasonally adjusting them might be problematic.
Most economic forecasting exercises involve variables with different seasonal patterns, and it is probably easier to work with seasonally adjusted data. You would have to embed the deseasonalisation within your model, which raises model complexity, and probably introduces extra errors that we know that we can explain.
If your objective is to forecast the non-seasonally adjusted variable, the usual practice in finanicial market economics is to come up with an estimate of the non-seasonally adjusted variable, then apply the existing seasonal pattern. The main example I have in mind is the forecasting of the CPI index in the inflation-linked markets, where the instruments are linked to the non-adjusted index. These forecasts were short-term (few month horizon), and were expected to be fairly accurate. Even long-term forecasts had a seasonal pattern built into them, since it mattered for relative value even for instruments with longer maturities.
Academics may prefer other techniques, but I am unaware of their use in short-term inflation forecasting.
If you are attemptng to forecast an aggregate index by forecasting its components, you might need to forecast each component, and then aggregate and seasonally adjust using the same algorithms used by the statistical agency. However, it’s rare that you can get the level of accuracy needed for this to matter.
answered Oct 14, 2017 at 12:15