statcast-era-pitches

This dataset contains every pitch thrown in Major League Baseball from 2015 through present, and is updated weekly.

Why

This data is available through the pybaseball pacakge and the baseballr package, however it is time consuming to re scrape statcast pitch level data each time you want / need to re run your code. This dataset solves this issue using github actions to automatically update itself each week throughout the baseball season. Reading the dataset directly from the huggingface url is way faster than rescraping your data each time.

Usage

With statcast_pitches package

pip install git+https://github.com/Jensen-holm/statcast-era-pitches.git

Example 1 w/ polars (suggested)

import statcast_pitches
import polars as pl

# load all pitches from 2015-present
pitches_lf = statcast_pitches.load()

# filter to get 2024 bat speed data
bat_speed_24_df = (pitches_lf
                    .filter(pl.col("game_date").dt.year() == 2024)
                    .select("bat_speed", "swing_length")
                    .collect())

Notes

• Because statcast_pitches.load() uses a LazyFrame, we can load it much faster and even perform operations on it before 'collecting' it into memory. If it were loaded as a DataFrame, this code would execute in ~30-60 seconds, instead it runs between 2-8 seconds.

Example 2 Duckdb

import statcast_pitches

# get bat tracking data from 2024
params = ("2024",)
query_2024_bat_speed = f"""
    SELECT bat_speed, swing_length
    FROM pitches
    WHERE 
        YEAR(game_date) =?
        AND bat_speed IS NOT NULL;
    """

if __name__ == "__main__":
    bat_speed_24_df = statcast_pitches.load(
        query=query_2024_bat_speed,
        params=params,
    ).collect()

    print(bat_speed_24_df.head(3))

output:

	bat_speed	swing_length
0	73.61710	6.92448
1	58.63812	7.56904
2	71.71226	6.46088

Notes:

• If no query is specified, all data from 2015-present will be loaded into a DataFrame.
• The table in your query MUST be called 'pitches', or it will fail.
• Since load() returns a LazyFrame, notice that I had to call pl.DataFrame.collect() before calling head()
• This is slower than the other polars approach, however sometimes using SQL is fun

With HuggingFace API

Pandas

import pandas as pd

df = pd.read_parquet("hf://datasets/Jensen-holm/statcast-era-pitches/data/statcast_era_pitches.parquet")

Polars

import polars as pl

df = pl.read_parquet('hf://datasets/Jensen-holm/statcast-era-pitches/data/statcast_era_pitches.parquet')

Duckdb

SELECT *
FROM 'hf://datasets/Jensen-holm/statcast-era-pitches/data/statcast_era_pitches.parquet';

HuggingFace Dataset

from datasets import load_dataset

ds = load_dataset("Jensen-holm/statcast-era-pitches")

Tidyverse

library(tidyverse)

statcast_pitches <- read_parquet(
    "https://huggingface.co/datasets/Jensen-holm/statcast-era-pitches/resolve/main/data/statcast_era_pitches.parquet"
)

see the dataset on HugingFace itself for more details.

Eager Benchmarking

Eager Load Time (s)	API
1421.103	pybaseball
26.899	polars
33.093	pandas
68.692	duckdb