Have a family member who is a mathematician and wrote a book on chaos theory. It's ironically about the predictability of chaos.
I don't really use any presently, fundamental events I've used in the past. But others might use TICK, TRIN or whichever breadth indicator.
The notion that TA "reductionism" is irrelevant. Sure one can show equivalency between indicators. Analogy is that all people are made of ratio of x pounds of chemicals and y water. It says nothing of the personalities nor the presentation of physical traits. The one thing it does tell you is the basic data, price, volume, is fundamental. Furthermore, indicators can be redundant. Hope that helps.
Become an expert in machine learning. Study the problem of market prediction deeply. Then you will see the pros and cons of using TA.
I have no clue about machine learning. I also cannot program (I use existing scripts and adapt the mathematical part in it to what I want to calculate. Blood, sweat and tears, but it works.) There are many ways to become a successful trader. There are experts in machine learning that never make any money as there are also wizards in math or PHD's that never make any money. So being an expert in any area is not equal to making money as a trader. If researching the wrong ideas , even the smartest researcher will never have any positive result. Reminds me of the (real?) story about the Russians using a simple pencil to write in space whereas the NASA spent millions to create a pen to write in space. Thinking out of the box and having a very analytical brain sometimes helps more than being an expert. There are many examples that proof that. Many discoveries are the result of researchers that accidentally do something they never did before what resulted in a new invention. https://www.visualcapitalist.com/accidental-inventions/
My input data is 1-minute OHLC. That data is further sliced, diced and organized according to my own logic and further divided into various parameters and statistics. So, while the raw data is 1-minute OHLC which I agree have its limitations - there's much more data made available from those simple data points than mere OHLC.