Every backtest is judged against the one path that actually happened. You can walk-forward, you can bootstrap, you can purge and embargo your CV folds, at the end of the day the strategy still only had to survive 2010–2023 in the exact order it occurred.. half of what looks like alpha is probably just path luck.

If you trained a generative model on returns and ran the backtest across thousands of plausible alternative histories, the path-dependent stuff would get exposed pretty fast, no? Anyone actually tried this, or is there a reason it doesn't work that I'm missing?