This project analyzes the factors associated with forecasting baseball game outcomes and provides key recommendations for success to industry leaders.
In the highly competitive and financially driven landscape of professional baseball, accurately predicting game outcomes can significantly impact team performance, fan engagement, and revenue generation. By leveraging historical game data our predictive modeling approach aims to provide teams and stakeholders with actionable insights into future game outcomes.
The dataset was taken from https://www.retrosheet.org/. They came from the Game logs dataset for 2010 through 2023. The data dictionary can be found on the webpage and in the supplemental materials of this repository.
After concatenating the thirteen years of data there were 32,484 regular season games (rows of data) and 179 columns/features.
The target 'win' was created using the (a) score of the visiting team, (b) score of the home team, (c) name of the visiting team, and (d) name of the home team to calculate the winning and losing team. The rows that represented each game were then divided into two row per game, one for each team and the winning team of the pair was assigned a win=1 and the losing team of the pair was assigned an win=0. Each pair was given the same ID and kept together through the train_test_split.
X features varied by model and included: 'at_bats', 'hits', 'double', 'triple', 'home_run', 'sacrifice_hit', 'sacrifice_fly', 'hit_by_pitch', 'walk', 'intent_walk', 'strikeout', 'stolen_base', 'caught_stealing', 'left_on_base', 'pitchers_used', 'wild_pitches', 'assists', 'errors', 'double_def', 'obp'. All features used in X were numeric and standard scaled.
This analysis spanned the COVID-19 pandemic. The pandemic introduced unique challenges for analyzing baseball data, requiring careful consideration of context, data quality, and potential biases when interpreting and modeling data from 2020 and beyond. In 2020 there were 60 regular season games per team, less than half of the standard 162.
Baseball has one of the richest data caches in sports, offering a plethora of statistics and metrics for analysis. Nonetheless, navigating this vast sea of information poses challenges, demanding a focused approach, organization, and an understanding of big data modeling techniques to extract meaningful insights from the wealth of available data.
Data are split into a training set and test set for modeling. Games are paired so that features for the winning team and losing team of a game are not split between the training and test set to prevent data leakage. Features are standard scale standardized. No additional preprocessing is used. There are no categorical features, no missing data, and the target is 50:50 balanced.
The primary evaluation metric used for these models is accuracy. With a balanced dataset and equal interest in classification of wins as much as losses, accuracy is an appropriate evaluation metric.
- The first model is a Dummy Classifier. As expected using the most frequent strategy the accuracy score of this balanced data set is 50%.
- The second model is a simple Decision Tree Classifier using parameters for criterion, maximum depth, and the full set of X features listed above. The accuracy score for this model is 75%. The model identifies: obp, left_on_base, intent_walk, at_bats, home_run, sacrifice_fly, and double as the most important features.
- Next a Random Forest Classifier uses a grid search for the estimator, and maxium depth and features. The best parameters are used in the model with the full set of X features listed above. The accuracy score for this model is 80%. The model identifies: obp, hits, at_bats, left_on_base, home_run, intent_walk, and errors as the seven most important features.
- The first Logistic Regression model uses the full set of X features listed above. The accuracy score for this model is 78%.
- The second Logistic Regression model uses the subset of X features identified as the most important features from the Random Forest: obp, hits, at_bats, left_on_base, home_run, intent_walk, and errors. The accuracy score for this model is 76%.
- The last model is a Neural Network. This model is built using extensive support from ChatGPT. The full set of X features listed above are used in the model. The accuracy score for this model is 82%; the best performing model. The seven most important feature identified in order of importance are: at_bats, left_on_base, walk, obp, pitchers_used, hits, and home_run.
While pitching and defense are crucial for well-rounded teams, the most important factors identified by this analysis when scouting is offense. Players who can get on base, especially those with a keen eye for spotting balls over strikes and earning walks, create more opportunities for additional at-bats.
