Complete agreement on that. Hence, "TA strategies" that are up to academic publishing standards usually look like the Bayesian regression method by Shah / Zhang, with trades being less than a minute apart.

So, no, absolutely no chance for me to show that my method produces statistically significant results based on the backtest results, way too few data points for that. But I said that much right from the beginning.

Guess it runs down to the question if the only methods (not just in TA) that you are willing to employ are those that can be fully rigorously shown to work, or if you allow including those that you have strong enough reasons to believe are working as well, but where those reasons can't be formalized to the point where you can show it in a "publishable" way. My example of choice in this context is chess opening theory, and the early (failed) attempts of chess algorithms beating human GMs practicing opening (or end game) theory.