Turning prediction data into product gears
I'm pursuing my Master's in Applied Machine Intelligence at Northeastern, with a first-author preprint on reinforcement learning in NCAA ice hockey, validated across 55 Division I programmes. My background spans business, management, and Human Resources, which shapes how I think about who uses a model and whether they can trust its output. I write here about the parts of a project that don't make it into a resume, the pipeline before the model, the validation step that's easy to skip under a deadline and focused on building toward user-centric solutions.
Markdown supported · ⌘↵ to post
Kartik N V J K
The leave-one-season-out call is the right instinct. With five seasons the failure mode is exactly what you named: sharp on training data, dead on a season it never saw. I would also slice-check the rare states, since an 8-state space can hide a class showing up only a few times. Did any state end up too sparse to trust?