🏀 NBA Fantasy Value Predictor (9-Cat)

Overview

Success in 9-category Head-to-Head (H2H) fantasy basketball requires more than just picking good scorers. It requires balancing nine distinct statistical categories (PTS, REB, AST, STL, BLK, 3PM, FG%, FT%, TO).

This project builds a Machine Learning pipeline to predict a player's future fantasy value (Z-score) based on their past performance. The goal is to identify undervalued players and support data-driven drafting decisions for the upcoming season.

Objectives

• Predict 9-Cat Value: Forecast the total Z-score for NBA players for the upcoming season.
• Beat the Baseline: Outperform the naive prediction method (assuming next season's stats = last season's stats).
• Feature Engineering: Analyze how rolling averages and age curves impact performance.

Tech Stack

• Language: Python 3.9
• Libraries: Pandas, NumPy, Scikit-Learn, Statsmodels, XGBoost, Matplotlib/Seaborn
• Data Source: NBA API & Basketball Reference (2020-2025 seasons)

Methodology

1. Data Collection & Cleaning

• Extracted per-game stats for ~2000 players across 5 NBA seasons.
• Normalized data using Z-scores relative to the league average for each season.
• Handled missing data and filtered for players with significant playing time (>50 career games between 2020-2025).

2. Feature Engineering

• Rolling Averages: Created 2-year rolling averages to capture recent form.
• Deltas: Calculated year-over-year changes to detect improvement/decline trends.
• Demographics: Included Age and Age² (quadratic term) to model the physical peak of athletes.

3. Modeling

I framed this as a regression problem, testing multiple models to predict the Next Season Total Z-Score.

Model	MAE (Mean Abs Error)	R² Score	Notes
OLS (Statsmodels)	1.750	0.714	Best Performer. Used backward elimination for feature selection.
Random Forest	1.816	0.692	Good at capturing non-linearities but slightly overfitted.
XGBoost	1.801	0.686	Robust, but slightly behind OLS in this specific dataset.
Baseline	1.857	0.669	Naive prediction (Next Year = Last Year).

Key Findings

1. OLS Supremacy: Surprisingly, the simple Ordinary Least Squares (OLS) model performed best. This suggests that the relationship between past and future stats is largely linear.
2. The "Scoring" Bias: Feature importance analysis showed that Points (PTS) and Points Per 36 were the strongest predictors of future fantasy success.
3. Predictability: Calculating the "stability" of stats showed that Assists and Rebounds are highly predictable year-over-year, while FG% and FT% are volatile and harder to forecast.

How to run

1. Clone the repo by - git clone https://github.com/maorshavit-boop/Predicting-fantatsy-basketball-value.git
2. Install dependencies by - pip install -r requirements.txt
3. Open the notebook in Jupyter Lab/Notebook

Project Structure

├── data/               
├── plots/              
├── code/               
├── requirements.txt    
├── .gitignore          
└── README.md