How I Spent 6 Months Building a Pitch-by-Pitch Baseball Simulator

Ten years ago, the author began building their first baseball simulator for a mere $2 on FanDuel. Frustrated by losses, they sought to understand why their picks underperformed and discovered the influence of luck and variance. Initially, their simulators were crude, using projections from RotoGrinder to generate skewed normal distributions for each player, selecting a single value from the curve for simulated scores. After 10,000 simulations, they solved for the optimal lineup based on these results.

Though pre-AI, this process eventually won three $100,000 prizes for a friend. The author's simulator lacked correlation between player performances and eventually inspired the creation of a pure play-by-play simulator. This was challenging both technically and physically, as the author had limited compute resources. They built an extensive back-testing suite against over 60 days of historical odds data before attempting to build the simulator, which served as the standard for its usefulness.

After the first successful back test, the author experienced a significant return on investment (+15%). However, they soon discovered data leakage, which caused a substantial drop in ROI (-22%). Despite this setback, the author persevered, showing unwavering dedication to creating the best simulator possible. They pushed themselves to design the system as if they had access to vast computing resources, even with a single computer.

This effort spanned months, with back-tests taking hours and fan noise keeping their family awake. Eventually, the author achieved their goal, completing a simulator that met their high standards.