Paulj1989 / expected-kills

Predicting binary outcomes (kill or error) in volleyball to produce an Expected Kills (xK) model.
MIT License
1 stars 0 forks source link

Examine model calibration #5

Open Paulj1989 opened 2 years ago

Paulj1989 commented 2 years ago

Is your feature request related to a problem? Please describe. Something I haven't looked at yet is how well the model is calibrated, and given that the probability of a kill is of particular interest in this case (much like the xG value is what is important, rather than the classification label), it seems like an important factor to consider how to improve the model's performance.

Describe the solution you'd like I need to explore methods for calibrating models, such as Platt's Scaling or I-Spline Smoothing. I think the latter will make more sense in this case.

Additional context Some resources for addressing calibration issues: