That's actually been surprising to me but makes sense: "the entire history of the stock market" is not that big in terms of data, and the dynamics change freque

That's actually been surprising to me but makes sense: "the entire history of the stock market" is not that big in terms of data, and the dynamics change frequently enough to where things that happened in 2000 are almost useless from a machine learning standpoint, my models perform best when trained on a much more selective subset (I believe because the reasons for trades and market participants keep changing, so the dynamics change enough to where old rules don't apply anymore).