Model Analysis Notes
Model Performance summarizes completed shot predictions and compares qualified model picks
against the raw baseline.
Top Picks Table provides the row-level player predictions, model-quality metrics,
shot distribution indicators, and final hit/miss results.
Model predictions are generated independently for each season. Hockey is highly temporal, and player performance can shift due to factors such as recent form, injuries, line combinations, roster changes, coaching strategies, and league-wide trends. A walk-forward approach beginning on November 1st allows the model to establish current-season patterns before predictions are evaluated.
Why We Use F1 Score
Overall accuracy can be misleading for NHL shot predictions because most players finish
below a given threshold. Instead, we evaluate our models using F1 Score, which balances
Precision (how often selected picks hit) and Recall (how many successful players
the model identifies).
By default, Top Shelf Analytics displays model segments with a minimum F1 Score of 75. This threshold balances pick quality and player coverage, helping keep the model focused on stronger predictions while still providing a meaningful pool of players across each slate.
Read the full article: Why We Use F1 Score Instead of Accuracy for NHL Shot Predictions