What is the reasonable accuracy?

I have read many research papers over the last couple of months. Many of these papers utilises talib, but claiming they were able to reach 90+ accuracy on binary classification.

I have spent the last 6 months in ML trading. The highest accuracy I was able to reach is 67% via xgbclassifier. My dataset has around 1 million rows.

I’m starting to wonder how these research papers were able to reach 90+ accuracy.

What is the highest accuracy you have achieved in ML trading binary classification? No need to provide your solution. I’m just trying to do a survey to understand what is possible.