Handcrafted keyword features (such as rating upgrades, guidance downgrades, buybacks, etc.) undergo sign flips between the training set and the test set; for example, "rating upgrade" was +0.28 before 2025 and became -0.04 after 2025.