backtesting with seasonally adjusted data

trapezoid

Newbie
Messages
3
Likes
0
Hi

Does anyone have any advice on seasonal adjustment in economic data and how it affects backtesting algos?

For example, if you take the 'latest' time-series of GDP, the seasonal adjustment smoothing applied to the data will mean a certain 'lookahead' in the data.

Do people normally ignore it, or only use some custom seasonal adjustment algorithm that is non-lookahead, or only do backtesting on non-seasonally adjusted data (which seems harder to find)?

thanks
 
It's not the seasonal adjustment that you need to worry about as that is reported in real time (meaning that figure is reported at the same time as the non-adjusted data). The concern is with the follow-on adjustments. GDP figures in the US, for example, have two scheduled revisions in the months following their initial release, never mind other revisions that could get made later.
 
Yes I know that several countries release their numbers in a few stages, and that is another problem that is hard to deal with - unless you have a full database of the original revision data.

Anyway, GDP is just one example, I'm interested in how generally to deal with seasonal adjustment which is used on many economic data series.

Although the seasonally adjusted data is reported real-time, I believe they typically re-calculate the seasonal adjustment factors and re-apply them to the whole historical data series.
So if you download the latest time-series, you know that the adjustment applied to the numbers 3 years ago was performed with knowledge of the most recent numbers.
 
You've basically made my general point - that being economic indicators are very often revised in some fashion or another after the fact - even well after the fact. Backtesting with economic data is only really valid if you can use the originally reported data, unless the adjustments are small enough that the would not have represented a meaningful difference in market reaction had they been reported at the time.
 
But how does one know if the 'adjustments are small enough' without the original revisions?

Yes any revisions would technically invalidate the backtest, but so what then? give up on the indicator?

I'm looking for practical advice from anyone who has dealt with this issue. Do they use non-seasonally adjusted data and then adjust using a backward-looking only filter? Or do they use non-seasonally adjusted data and use lookbacks that are factors of 1 year, so seasonality does not matter.

Or use seasonally adjusted data but with only long lookbacks , so any revisions have a small effect on the trading signal?

etc..
 
If you can get clean (unrevised) nominal data (non-seasonally adjusted), and really feel the need to have the seasonal adjustments, one thing to consider is to do you own adjustments - obviously only using the data you would have had available to you at any given point in time. This is something a neural net application (for example) might prove very handy for.
 
Backtesting using fundamental data is interesting but in most countries the numbers do not reflect economic reality.
 
Top