Thanks for writing this piece. I have had the same issues however I have gained some improvement in performance using cryptocurrencies. However, I do label my data differently. To really tackle this problem you would need a super computer. Just imagine if you pick one company, 10 time bars, one algorithm, label the data 100 different ways, try several feature selection schemes (10), and imagine this model takes 10 seconds to train (very optimistic). This would amount to

10 types of time bars*1 model*1 algorithm*100 labeling methods*10 feature selection schemes*10 seconds to train = 100,000 seconds

That’s just for one market for just one algorithm. Imagine if you wanted to explore 10 algorithm, 5 different types of bars, and 100 markets, and different types of labeling. Hence why you really need a super computer to really tackle this problem. However, I think this is manageable if you use some tricks to reduce the computation time.