Model Card: NBA Fantasy App
Model Details
Developers: Hatman Model Type: Functional Keras Model Language: Python Training Data: 2018 NBA season game results and player stats Evaluation Data: Same as training data Intended Use: Predicting winners of simulated NBA games based on player stats Limitations: Does not account for changes in playstyle, three-point line, or other rule/environment changes over time.
Model Description
This model is a functional Keras neural network trained on the 2018 NBA season's game results and player statistics. The model takes the season averages of various box score stats (points, rebounds, assists, etc.) for the starting 5 players on each team as input features. The target is a binary classification of whether the team won or lost the game.
The model architecture consists of:
- Input layer with 100 nodes (5 players x 10 stats per player)
- Two dense layers with 50 and 64 nodes respectively, using ReLU activation
- Output dense layer with 1 node and sigmoid activation for binary classification
The model was trained using the skLearn's RandomizedSearchCV with loss function BinaryCrossentropy and the RMSprop optimizer. Training script included.
Intended Use
The primary intended use of this model is to predict game winners in the NBA Fantasy App, which allows users to create fantasy teams from historical NBA players and simulate games against computer teams. The model provides win/loss predictions based solely on the stats of the starting lineups.
Factors for Predictions
The key factors used in the model's predictions are the season averages for various box score statistics of the starting 5 players on each team, such as:
- Points
- Rebounds
- Assists
- Steals
- Blocks
- Field Goal %
- 3-Point %
- Free Throw %
Caveats and Recommendations
While reasonably accurate on 2018 season data, this model has some important limitations:
Data Recency: It was trained only on the 2018 season, so its predictions may be less reliable for other seasons with different playing styles, rules, etc.
Player Aging: The model does not account for the aging/development of players over time. A player's stats from 10 years ago may not reflect their current abilities.
Team Dynamics: It considers only individual stats and not team-level factors like coaching, chemistry, or system fit.
Rule Changes: The model does not adapt to changes in rules that could impact team strategies and player values (e.g. moving the 3-point line).
It is recommended to use this model's predictions as one factor among many when evaluating potential fantasy teams, rather than treating them as definitive outputs. Incorporating additional contextual information about teams, players, and rules could improve prediction accuracy.
For production use cases, it is advised to retrain the model regularly on the most recent NBA data to account for changes in the game over time.