Our general philosophy is to start as simple as possible and meticulously and methodically investigate features, engineered features, and feature importances one-by-one.

The Progression:

• Model 0 (Baseline): The simplest possible model would assume a team scores their average points per game, and the opponent scores their average. This fails to account for opponent strength or pace of play.
• Model 1 (Simple Average): Our simplest viable model. We estimate pace and efficiency by simply averaging the values of each team. This forms our physics-based baseline.
• Models 2-4 (Variance Investigation): We hypothesized that accounting for variance would improve accuracy. For example, if a low-variance team like Houston (62 ± 5 possessions) plays a high-variance team like USC (86 ± 16 possessions), we might want to weight the expected pace towards Houston. However, our analysis showed that the simple average (Model 1) actually performed better than inverse variance weighted terms.
• Model 5 (Linear Regression): A linear regressor using only the Pace * Efficiency term. This validates that a learned coefficient on the fundamental scoring equation performs at least as well as the simple multiplication.
• Models 6-7 (Feature Engineering): We methodically researched and engineered a large set of features. For Phase 1, we brought forward the most "vanilla" of these features (e.g., efficiency, pace, team strength, venue) to build our XGBoost models.

Phase 2: We have accomplished our Phase 1 goal: a framework and pipeline for researching, validating, and deploying ML models. We are now deep into Phase 2, involving data-driven feature engineering, new objective functions, and more advanced model architectures and ensembling.