How far back does your data actually need to go?

Both are real. Longer gives you more regimes, which is good for robustness and bad for relevance. I look at the long period for "does it survive different weather" and the recent period for "is it working now."

And when those two disagree, that disagreement is itself the most useful result you'll get.

Depends a lot on your timeframe too. I trade 5m and honestly intraday data from a decade back is a different animal, session behaviour, who's on the other side, how fast gaps get filled. For me the long history is mostly a sanity check that the setup isn't a one-regime fluke, and the last couple of years is what I actually tune on. Also worth checking your data quality that far back before you draw conclusions from it, old intraday bars can be messier than people assume.