Has anyone taken a course that moves beyond descriptive charts into building predictive models for decision-making? I’m piloting a 6-week module where coaches calculate serve+1 expected value from 2024 hard-court data using Point-By-Point CSVs and Google Sheets/BigQuery, and I’d like recommendations for CE programs that cover model validation and practical thresholds (e.g., when a 3% EV gain justifies a pattern change).
I’d look at Google Cloud’s Machine Learning with BigQuery on Coursera (https://www.coursera.org/learn/machine-learning-bigquery); it covers validation, calibration, and thresholding in the exact stack you’re using. For serve+1 EV, use a time-based split by tournament week to prevent leakage and choose the threshold by maximizing expected points gained after isotonic calibration in BQML — think of it like string tension: too tight and you spray, too loose and you float. Also, are you logging per-player calibration so you can set different thresholds for servers vs returners; that usually stabilizes week to week.
Building on @ethanR00: run a reliability plot + Brier score in Sheets; set ‘serve+1’ threshold where EV gain >0.02.
Quick win: split by match date with a forward/rolling setup (train earlier months, validate next block, test the last) and pick your action threshold by maximizing net win‑probability added using point‑importance weights — no future peeking, like switching to new balls each set. For CE that goes deep on validation and thresholding, University of Washington’s Practical Predictive Analytics: Models and Methods is solid: https://www.coursera.org/learn/predictive-analytics. @liam_sanchez45 have you tried leverage‑weighted calibration here?