How do I create features that actually improve a model?
You can use the features handed to you, and creating new ones that genuinely help feels out of reach. That's a normal place to be. Using features is following a recipe. Creating them is understanding why the dish works. The second skill grows out of the first, with time. Good features come from understanding the problem, not from tricks. When you really know what drives the thing you're predicting, the useful features start to suggest themselves. It's less about math and more about asking, what actually matters here, and how could I measure it. Curiosity about the domain is the real engine. Take one problem you know well and try to invent a single new feature from your own understanding of it, then test whether it helps. One honest attempt teaches more than reading about it. Creating features isn't beyond you. It's a skill you start building the first time you try.