About

Bench Coach exists because the public-facing MLB analytics tools were loud and thin or rigorous and unreadable. Neither showed the math. Neither treated the reader as someone who could follow it. There was room for a tool that did both.

The model started as a linear algebra observation: Markov chains can model state-dependent outcomes — systems where the next state depends only on the current one. Baseball fits the framework more cleanly than most sports — twenty-four discrete game states, clean transitions, outs as the absorbing state. Retrosheet's 2.6 million plate appearances across 2010–2024 made the transition matrix estimable to three decimal places. The 2017–2025 backtest set gave the calibration check — Brier 0.1598–0.1677 across 9 out-of-sample seasons (361,519 inning-start predictions), 82.1–84.8% accuracy at ≥70% confidence, spanning four MLB rule eras. Bench Coach is the productized version — FastAPI and Next.js on top of the math, a live GUMBO feed for state, and a calibration curve that has to hold up against actual sportsbook lines.

The scorebook aesthetic is its own argument: numbers deserve typography. Playfair Display for headers, Source Serif 4 for body, JetBrains Mono for everything that counts.

Opening Day 2026 was the go-live. The model is calibrated, the dashboard is quiet, and the math is visible. Bench Coach is the product of human-AI integration.

The Platform

Bench Coach

A real-time MLB betting analytics platform. 75-state lead-aware Markov chain, seven-input win-probability model, 5,000 Monte Carlo simulations per query, live sportsbook comparison, and backtested calibration. See the full methodology.

Contact