Well generally speaking, we all enter a strategy under at least one assumption, that future price will belong to the same distribution family as past price. So there is always an advantage to testing over longer periods since the future price might take on characteristics of some older version of the distribution. On the other hand, if you are optimizing your parameters, then you might want to focus on more local data to catch subtle changes which can have a huge effect on performance. So the way I think about it, back-test over as much time as possible, and optimize over as little time as possible. (of course the optimization window should be tuned as well.)- Would recent data be exponentially more import?
- How likely is it that systems are being modified unfavorably to suit conditions not relevant to today?
- What is the expected life of an automated system before it requires tweaking?
Obviously performance needs to be constantly monitored to spot weaknesses...and the system needs a real chance to prove itself without constant tweaking but It makes sense to me at least that recent data should be valued higher than data from 2020.