Capstone Project: Using a Batter's Offensive Statistics to Predict Their Last Season Played Batting Average, Runs Batted-In (RBI's) and Homeruns

Author: Mario Sanchez, Jr.

Date: 8/25/2019

Question:

Can regression models be created to accurately predict a players batting average, runs batted-in (RBI's) and homeruns of their last season played?

Objectives:

• Obtain baseball data from trusted sources.
• Learn how to aggregate the multiple csv files into the form I need.
• Pick multiple regression models and compare their performance on unseen data.
• Which model performed best for each metric (HR's, RBI's, AVE)?
• Determine which features are most important to predicting each target.
• How can we use this information to improve our ability to predict a players performance using historical data?

Data Description:

• Batting.csv
• People.csv
• retrosplits
• Most of the data in the Databank is provided by Chadwick Baseball Bureau (http://www.chadwick-bureau.com). The data differ from the data the Bureau provides to its clients in that it contains less detail, is updated less frequently, and is provided on an as-is basis.

The Databank is historically based in part on the Lahman Baseball Database, version 2015-01-24, which is Copyright (C) 1996-2015 by Sean Lahman.

The tables Parks.csv and HomeGames.csv are based on the game logs and park code table published by Retrosheet. This information is available free of charge from and is copyrighted by Retrosheet. Interested parties may contact Retrosheet at http://www.retrosheet.org.

• batter_and_change_FINAL DataFrame has 41,745 rows and 84 columns
• previous_years_FINAL DataFrame has 38,908 rows and 84 columns
• last_year_df_FINAL DataFrame has 2837 rows and 84 columns

Data Dictionary:

Columns Name	Description
playerID	unique identifier
yearID	year for that row
teamID	team played on
stint	stint
G	games played
AB	at-bats
R	runs
H	hits
2B	doubles
3B	triples
HR	homeruns
RBI	runs batted in
SB	stolen bases
CS	caught stealing
BB	base on balls
SO	strike out
IBB	intentional walk
HBP	hit by pitch
SH	sacrifice hit
SF	sacrifice fly
GIDP	grounded into double plays
nameFirst	first name
nameLast	last name
bats	left or right
throws	left or right
debut	first year played
finalGame	last year played
_percent	percent spent in that era
era_	binary era
decade_	binary decade
throws_R	1 if throws R
bats_R	1 if bats R
AVE	average
OBP	on base percentage
Slug_Percent	slugging percentage
OPS	on base + slugging
debutYear	first year played
currentYear	year of that row
YRSPRO	experience
_chg	change from previous year
KMeans_label	cluster label

Sources:

1. https://medium.com/@williamkoehrsen/data-analysis-with-python-19434f5d6324
2. http://www.seanlahman.com/baseball-archive/statistics/
3. https://www.datacamp.com/community/tutorials/scikit-learn-tutorial-baseball-1
4. https://www.datacamp.com/community/tutorials/scikit-learn-tutorial-baseball-2
5. https://github.com/WillKoehrsen/Udacity-Data-Analyst-Nanodegree
6. http://www.baseballdatascience.com/

Target: Last Year Played AVE, RBI's and HR's

EDA and Model Selection:

1. Obtained historical baseball data from many online resources.

2. After exploring many of the files available I decided to use the one which contained each players stats divided by years.

3. Combined the batting.csv with people.csv to obtain vital information on each player.

4. Grouped my data by player, year and team.

5. Filtered my dataframe to only include batters who had at least 20 games played and 20 ab-bats.

6. Next I wrote a function to include only players who had at least 5 years played in the MLB.

7. I then filtered my dataframe one last time to include only players who appeared after 1900.

8. Feature Engineering:

• Wrote a function to assign an era label to each player dependent on when they played the game and then what percentage of their career they played in that era.

• Assigned a year label to each player and created dummy columns from these labels.

• Assigned a decade label to indicate which decades the player had played in.

• The above columns were created to account for the different era's that have occurred over the last 119 years.

• Created a binary column to indicate if a player batted and threw right handed.

• Created AVE, OBP, Slug_Percent, and OPS columns.

• Computed a players experience by subtracting the current year from the debut year.

9. Split my final dataframe into previous years and final year so that I can test my models on unseen data and compare their results.

10. Model Preparation:

• Used previous years to train and test on
• Made sure that columns such as G, H, AB were left out
• Scaled my train, test and unseen data for some of the regression models
• Created polynomial features to my X variable to provide more data to my models
• Grid search over several parameters for each model
• Created a pickle file of each fit and trained model for future evaluation.

11. Models:

• For each target, HR's, RBI's, and AVE, I fit each of the following models:
• Linear Regression
• Ridge
• Lasso
• ElasticNet
• RandomForest

Model Performance:

After many different iterations, the following models were tested with above features to determine which performed best using the R2 score and RMSE.

1. Predicting AVE:

• Linear Regression Train R2 Score: 0.927265749628238
• Linear Regression Test R2 Score: 0.9245850874144554
• Ridge Train R2 Score: 0.9641036826810402
• Ridge Test R2 Score: 0.9501996596045186
• Lasso Train R2 Score: 0.9167955525718006
• Lasso Test R2 Score: 0.9140641297211587
• ElasticNet Train R2 Score: 0.9471689658637862
• ElasticNet Test R2 Score: 0.9453456477932288
• Random Forest Train R2 Score: 0.9935941144091776
• Random Forest Test R2 Score: 0.9543081856108968
• Linear Regression RMSE on New Data: 0.022072780140648625
• Ridge RMSE on New Data: 0.0323865503500996
• Lasso RMSE on New Data: 0.03298153698364595
• ElasticNet RMSE on New Data: 0.030976570580385922
• Random Forest RMSE on New Data: 0.016747452199047858
• Linear Regression R2 Score on New Data: 0.8157356416523764
• Ridge R2 Score on New Data: 0.6033050721934583
• Lasso R2 Score on New Data: 0.5885954929121789
• ElasticNet R2 Score on New Data: 0.6370941785539136
• Random Forest R2 Score on New Data: 0.893922137971072

2. Predicting RBI's:

• Linear Regression Train R2 Score: 0.9282337759628065
• Linear Regression Test R2 Score: 0.9311742490423233
• Ridge Train R2 Score: 0.9493429034345244
• Ridge Test R2 Score: 0.9430190274928
• Lasso Train R2 Score: 0.94951401869399
• Lasso Test R2 Score: 0.9465989701115267
• ElasticNet Train R2 Score: 0.9218334693805789
• ElasticNet Test R2 Score: 0.9250505846389491
• Random Forest Train R2 Score: 0.9887318112663641
• Random Forest Test R2 Score: 0.9362520045126567
• Linear Regression RMSE on New Data: 5.526281075333878
• Ridge RMSE on New Data: 26.189041645368615
• Lasso RMSE on New Data: 25.380538043708153
• ElasticNet RMSE on New Data: 24.269181261146084
• Random Forest RMSE on New Data: 5.213557078014037
• Linear Regression R2 Score on New Data: 0.8801086958714287
• Ridge R2 Score on New Data: -1.6925325162640654
• Lasso R2 Score on New Data: -1.5288518836501495
• ElasticNet R2 Score on New Data: -1.3122351284302
• Random Forest R2 Score on New Data: 0.8932937127367198

3. Predicting HR's:

• Linear Regression Train R2 Score: 0.9123830309085669
• Linear Regression Test R2 Score: 0.911395154652938
• Ridge Train R2 Score: 0.9839560929298725
• Ridge Test R2 Score: 0.9810375920991171
• Lasso Train R2 Score: 0.9895424128271637
• Lasso Test R2 Score: 0.9873949398289043
• ElasticNet Train R2 Score: 0.8732944355681218
• ElasticNet Test R2 Score: 0.873122027197329
• Random Forest Train R2 Score: 0.9914084916509395
• Random Forest Test R2 Score: 0.9565788564133643
• Linear Regression RMSE on New Data: 2.1945386661296733
• Ridge RMSE on New Data: 6.871186887441323
• Lasso RMSE on New Data: 6.6806050835395405
• ElasticNet RMSE on New Data: 5.7251772746060565
• Random Forest RMSE on New Data: 0.9997683023791144
• Linear Regression R2 Score on New Data: 0.6736586218690606
• Ridge R2 Score on New Data: -2.199257455812315
• Lasso R2 Score on New Data: -2.024247067469625
• ElasticNet R2 Score on New Data: -1.221076649274769
• Random Forest R2 Score on New Data: 0.932269482244308

Primary findings/conclusions/recommendations:

• The RandomForest Regression model performed the best on predicting all metrics.
• Many of the models performed well on test data but poorly on unseen data.
• This indicated to me that using an ensemble model was a better approach at more accurately predicting my targets.
• I believe that I can achieve even better results if I log transform my target variable since there was a skewed distribution for two of the three targets.
• The information obtained from the models such as the most important features can then be leveraged to direct scouting reports and help teams better evaluate their players performance. These models allow for a players past history as well as others from around the league to determine what type of batter they are.
• This can then allow for a better forecast of a team's performance broken down by player.

Limitations and Assumptions:

• More feature engineering can improve the models.
• It is possible that there may have been some data leakage, further investigation is needed.
• Computing power becomes an issue when fitting certain models. This greatly influences how much tuning can go into each model.
• The results are encouraging and I believe they can by extended to predict many other offensive metrics.

Future Analysis:

• Explore other models such as ExtraTreesRegressor, AdaBoostRegressor and BaggingRegressor.
• Test models on other offensive metrics such as on-base percentage (OBP), or Slugging Percent and see if they can generalize well to these metrics.
• Test the models on select subsets of players such as by position or by era. This may reveal very interesting findings.
• There is truly a mountain of available data for baseball as well as other sports and am excited to apply what I have learned in this project to future projects.