Neural Networks for Longitudinal Data Analysis
summary
Uses the Lahman baseball tables to predict the three-part slash line (AVG, OBP, SLG) from hitters with at least four consecutive seasons, building rolling four-season windows with padding and downsampling to limit long-career bias. Compares OLS (stats), GBM (h2o) and recurrent nets implemented with tensorflow and keras in R, tuning GRU/LSTM layers and dense heads; neural nets on a Colab GPU show much closer predicted vs actual, correctly separating Willie Mays, Greg Nettles and Chris Davis. Notes remaining issues: overlapping-window dependence, missing prediction intervals, and output-shape changes for multi-step forecasting; code and slides live in a GitHub repo.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:05What is longitudinal data
- 3:13Baseball data example
- 5:40Modeling for prediction
- 8:10Recurrent neural networks
- 10:40Data prep and windows
- 12:24Baseline and NN setup
- 14:47Implementation in R/Colab
- 15:09Model performance
- 17:18Limitations and future
- 18:54Q&A
related talks
56mSports Analytics: Big Data in the Big Time
38mDeep Learning plays Handball
30mApplying The Magic Of Neural Networks
01h 26mApplying Statistical Modeling & Machine Learning to Perform Time-Series Forecasting
31mThe Advantages of Deep Learning with Recurrent Neural Networks
39mForecasting With Classical and Machine Learning Methods
42mHow Deep Learning is Revolutionizing Sports Analytics
48mTBA