All market opportunities which can be predicted more often than a coin toss, are exploited away because people act on them now.
All that remains, to be made money on, are market opportunities which can be predicted less often than a coin toss but have a positive expected value, i.e. it lets you make more the times you are right than lose the times you are wrong.
This is why an 8-year-old can be right more often about the market than a professional, but only the professional can make money, not the kid because the when the kid is wrong more money is lost than what is made when the kid is right. When the professional is wrong, the professional loses less money than what is made when he/she is right.
There are many strategies that have positive expected value but demand different levels of blood. People with shallow pockets will be lined up around strategies that are right >40% of the time, but makes 3x and loses 2x. People will deeper pockets (e.g. Soros/Buffet/Banks) will be lined up around strategies that are right >10% of the time, but makes 9x and loses 1x.
The problem with using machine learning in the markets is that classifiers and regressors are trained to maximize win-rates not expected value or align with the “pocket deepness” of the trader. Another problem is that all (?) technical analysis indicators ceases to function when volatility changes. Moving average crossovers, crossover in high volatility even when there is no trend, Oscillators give signals in low volatility even when there is no mean reversion. So training on technical analysis indicators without using volatility as a hyper-parameter of the indicator parameters is a sure-fire way to lose money. Also if the training set or backtest data has a small uptrend, a strategy that has, by design, a tendency to go long will make more money even though it is driven by an unfair coin toss instead of a classifier.