How far back do you back test?
retained text
I have just recently been getting into this. I have started and scrapped several models so far, each time further refining my rules for constructing a model to limit bias and overfitting. In one of the legs of research I did I learned that the market fundamentally changed around 1999-2000, with evidence strongly suggesting we are in the process of another fundamental change, or that one has already occurred (probably either a little before or after covid). Though this wont be able to be confirmed until we have a few more years of data. But after discovering this, I changed how I was looking at the data. Yes, you want a robust system that is not just tuned to do well in the current market conditions. But at the same time, tuning for a market that no longer exists wont do you any good either. So clearly the answer is back testing only up to a certain number of year, and replacing the oldest year with the newest year on a rotation. But the question is, how far back do you go? I think ideally, I would like to focus on the past 25 years with a focus/weighted effect on the newest 5-10 years. So, how far back do you typically back test?
document fetched 2026-09-22 00:44 · canonical https://reddit.com/r/algotrading/comments/1wm7hzx/how_far_back_do_you_back_test