I’ve stopped manually triggering model training. Previously, I’d run training jobs reactively, usually after debugging something, or when I “felt” the data had changed enough. Not exactly scientific. Now, training happens nightly. Each target (bull, bear, range) gets its own dataset window and label schema. Models are versioned, and logs are dumped to an S3-equivalent so I can diff SHAP outputs or validation curves over time. No more “just one more tweak” loops. It either works or it doesn’t. Move on. [[ML]] [[Serendipity]]