⚾ Swing-and-Miss Probability Predictor

DSC 148 – Course Project | UCSD

Predicting whether a pitch results in a swinging strike using Statcast data from Baseball Savant.

---

Project Structure

swing_miss_project/
├── data/                    # Raw and processed data (git-ignored)
│   ├── raw/
│   └── processed/
├── src/
│   ├── data_loader.py       # Statcast API fetching & caching
│   ├── feature_engineering.py
│   ├── models.py            # LR, RF, XGBoost, Neural Net
│   ├── evaluate.py          # Metrics & plotting
│   └── utils.py
├── notebooks/
│   └── eda.ipynb            # Exploratory Data Analysis
├── results/                 # Saved figures & metrics
├── tests/
│   └── test_features.py
├── train.py                 # Main training script
├── predict.py               # Inference on new pitches
├── requirements.txt
└── README.md

---

Setup

# 1. Clone & enter the repo
git clone https://github.com/<your-username>/swing-miss-predictor.git
cd swing-miss-predictor

# 2. Create a virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

---

Data

Data is pulled automatically from Baseball Savant via the pybaseball library, which wraps the Statcast API.

# Fetch data for a season range (downloads ~500k+ pitches)
python src/data_loader.py --start 2022-04-01 --end 2023-10-01

Data is cached locally in data/raw/ to avoid re-downloading.

---

Training

# Train all models and save results
python train.py

# Train a specific model
python train.py --model xgboost

# Options: logistic, random_forest, xgboost, neural_net, all

---

Prediction

# Predict on a single pitch (example)
python predict.py \
  --velocity 94.2 \
  --spin_rate 2400 \
  --extension 6.2 \
  --release_x -1.5 \
  --release_z 5.8 \
  --plate_x 0.3 \
  --plate_z 2.1 \
  --balls 1 \
  --strikes 2 \
  --stand R \
  --p_throws R \
  --pitch_type FF

---

Models

Model	Description
Logistic Regression	Linear baseline; interpretable coefficients
Random Forest	Ensemble; captures non-linear interactions
XGBoost	Gradient boosting; typically best tabular performance
Neural Network	MLP with batch norm; learns complex pitch tunneling

---

Evaluation

All models are compared on a held-out 2023 test set using:

• Accuracy – overall correctness
• F1 Score – balances precision/recall (important: class imbalance ~10% swinging strikes)
• ROC-AUC – probability calibration quality

Results and plots saved to results/.

---

Key Features

Feature	Description
velo_diff_from_avg	Velocity relative to pitcher's season average
zone_*	Binary zone indicators (Statcast 9-zone + balls)
tunnel_dist	3D distance to previous pitch at decision point
prev_pitch_type	Encoded type of pitch thrown before
prev_pitch_result	Whether previous pitch was called/swung/missed
count_leverage	Custom count-pressure index

---

Results Summary

Results are generated after running train.py. A results/comparison.png chart is produced comparing all four models.