Mariners Player Development Research

An exploratory baseball analytics project examining whether early-season Statcast contact metrics can help describe and predict later-season contact quality for Seattle Mariners hitters.

Key finding: In this exploratory sample, a simple early-season Hard-Hit% model outperformed a more feature-rich model when predicting later-season Hard-Hit%, highlighting the importance of model validation and sample size in player-development research.

Research Question

Can early-season contact quality metrics provide useful information about a hitter's later-season performance?

The project focuses on:

• Exit velocity
• Hard-hit rate
• Barrel rate
• Player-level changes from early to late season
• Predictive modeling of later-season hard-hit rate

Data

The analysis uses publicly available MLB Statcast data for the 2026 Seattle Mariners season through August 28, 2026.

The collected dataset contains:

• 5,077 batted balls
• 106 unique game dates
• 119 Statcast variables
• 12 hitters meeting the minimum sample requirements for the predictive analysis

Raw Statcast data is intentionally excluded from the repository and can be regenerated using the ingestion script.

Methodology

The project is organized into several steps:

1. Collect Mariners schedule and Statcast data.
2. Filter Statcast observations to batted balls.
3. Aggregate contact-quality metrics at the player level.
4. Split the season into early and late periods using July 1 as the cutoff.
5. Compare changes in exit velocity, hard-hit rate, and barrel rate.
6. Build a simple linear regression model to predict later-season hard-hit rate.
7. Evaluate predictions using leave-one-player-out cross-validation.
8. Compare a simple model against an expanded model using additional Statcast features.

Player Development Analysis

Player-level changes were calculated by comparing early- and late-season contact quality.

The analysis identified different development patterns across hitters. Some players improved across multiple contact-quality measures, while others showed declines or mixed results.

For example:

• Mitch Garver showed improvement in average exit velocity, hard-hit rate, and barrel rate during his time with Seattle.
• Cal Raleigh improved across all three contact-quality measures.
• Dominic Canzone showed one of the largest declines across the measures examined.
• Brendan Donovan showed a mixed pattern, with average exit velocity increasing while hard-hit rate and barrel rate declined.

These classifications are exploratory and should not be interpreted as formal player-development grades.

Predictive Modeling

Initial Model

The first model used:

• Early-season average exit velocity
• Early-season hard-hit rate
• Early-season barrel rate
• Early-season average bat speed

The initial in-sample results were:

• R²: 0.650
• MAE: 3.2 percentage points

However, this evaluation used the same players for training and evaluation, which can overstate predictive performance.

Out-of-Sample Validation

To obtain a more realistic estimate of predictive performance, leave-one-player-out cross-validation was used.

The resulting performance was:

• R²: -0.288
• MAE: 6.1 percentage points

The negative R² indicates that the model performed worse than a baseline that simply predicts the average late-season hard-hit rate.

This demonstrates the importance of evaluating predictive models on observations that were not used during training.

Model Comparison

Two models were compared using the same leave-one-player-out validation approach.

Model	R²	MAE
Early Hard-Hit% only	0.210	4.74 pts
EV + Hard-Hit% + Barrel%	-0.053	5.34 pts

In this sample, the simpler model performed better than the expanded model.

This suggests that adding additional contact-quality variables did not improve out-of-sample prediction for this dataset.

Visualization

The visualization shows the change in average exit velocity and hard-hit rate from the early to late season for each hitter.

Limitations

This is an exploratory research project and should not be interpreted as a production player-development model.

Important limitations include:

• The predictive analysis contains only 12 players.
• The small sample limits the reliability and generalizability of the model.
• The barrel definition used in this project is a simplified exploratory definition.
• Public Statcast data does not include proprietary club-level information such as internal Trackman, Hawkeye, or organizational datasets.
• Player context, injuries, playing time, and other factors are not fully captured by these contact-quality metrics.
• The analysis describes observed relationships and does not establish causation.

Project Structure

MarinersPlayerDevelopmentResearch/
├── data/
│   ├── raw/
│   └── processed/
├── notebooks/
├── src/
│   ├── ingest.py
│   ├── features.py
│   ├── analysis.py
│   ├── visualization.py
│   ├── predictive_model.py
│   ├── model_comparison.py
│   └── player_development.py
├── .gitignore
├── README.md
└── requirements.txt