Learning Contextual Event Embeddings to Predict Player Performance in the MLB

This repository contains the code and (links to) data required to reproduce the experiments described in the abstract/paper "Learning Contextual Event Embeddings to Predict Player Performance in the MLB."

Our models are able to make pitcher single-game strikeout and batter has-hit predictions that are competitive with three major sportsbooks in the US. A comparison of the predictions is presented below. It is important to note the following:

1. Our model uses only 10 games worth of play-by-play data to make predictions
2. Our batter has-hit predictions reflect the batter recording a hit off of the starting pitcher only, while the books' predictions reflect a batter recording a hit during the entire game, arguably an easier task

Preliminaries

Before models can be trained, the data must be downloaded.

1. Download at-bat sequence files (ab_seqs_v17_ssac.tar.gz) here
   1. Extract the data, and remember the location of the ab_seqs_v17_ssac as it will be needed later.

Pre-training

To train the model described in the paper, simply execute the run_modeling.sh script found in the scripts folder. The model will take approximately 24-30 hours to train with the given parameters.

• Note: before running the script, please change the AB_DATA variable to match the location of the ab_seqs_v17_ssac directory mentioned above.

Fine-tuning to Predict Performance

To fine-tune the pre-trained model for performance prediction, simply execute the run_finetune.sh script found in the scripts folder. Fine-tuning will take approximately 4 hours with the given parameters. If you wish to use a different pre-trained model, modify the filepath given as the --model_ckpt argument.