To expand on what the other posters have said, humans are natural storytellers. We want things to have a narrative, and we are bad at determining if a theory merely fills our desire for a narrative. Starting from a known phenomenon and crafting a theory around what we know should happen has the danger of creating a model that is too specialized or completely contrived.

If, however, the theory is able to make predictions about things we didn't already know, and those predictions agree with experimental data, then there's a higher chance our new model maps to physical reality.