• Metadata

  • Source: https://github.com/elsehow/aaronson-oracle
  • Repo: https://github.com/paulliwali/aaronson-oracle-baseball

• Adapting Aaronson’s Oracle to predict baseball pitches for a complete game

• Aaronson’s Oracle

  • Algorithm that predicts the next keystroke by analyzing the past 5 keystrokes and the 32 possible combinations they can create
  • Then see what is more likely to follow that 5 keystroke combination
  • So when the user starts to type a particular sequence it will guess what the next keystroke is
  • It is 70% accurate
  • Enhancements can be also storing other various lengths and use the most confident one

• Pseudo-code

  • initialize a dict with 5-gram combinations as keys and the number of times each keystroke follow this combination as dict of values
  • record the rolling 5-gram combination from the user
  • look up the highest probable next keystroke, return that or default to one by chance if there are no stored values
  • update the dict with user inputs
  • calculate the rolling accuracy when making a prediction

2026-01-01: performance

The existing performance of the models were:

Model	Overall	Fast	Breaking	Off-Speed	N Pitches
Naive (Always Fast)	0.5643	1.0000	0.0000	0.0000	24,917
N-Gram (n=3)	0.5335	0.7795	0.2496	0.1121	24,917
N-Gram (n=4)	0.5376	0.8223	0.1977	0.0837	24,917
Frequency-Based (Oracle)	0.4834	0.5979	0.3775	0.2104	24,917
Markov Context	0.5552	0.9493	0.0540	0.0175	24,917

2026-03-15: autoresearch

Leveraging Andrej Karthpath’s autoresearch loop to discover better models for predicting pitch sequencing. After a night of running and my Macbook dying half way, it was able to run 35 experiments and discover a model with 0.61% accuracy on the evaluation dataset that had an architecture of 6-layer transformer, 4 heads, 128d model, FFN 6x hidden dim with some regularization. It also found that other architectures like GBM or LSTM performed much worse.

Side note on how it works -> autoresearch

#data-science #projects