Model v2.1Accuracy: 55.7% (472 live)Brier: 0.2468Live:55.7%472/500
Today: 15 picks 7 value playsAPI

About Diamond AI

Not an AI Guessing.
A Trained Statistical Model.

Diamond AI is a quantitative MLB prediction engine. At its core is a logistic regression trained on 22,762 historical games, validated at 56.5% accuracy on a held-out season. Every prediction is driven by math, not opinion.

56.5%

Validated Accuracy

2,430 holdout games

22,762

Training Games

2015–2024 seasons

233K+

Historical Games

Retrosheet 1871–2025

0.243

Brier Score

0.25 = coin flip

The Prediction Stack

Four layers, each adding measurable accuracy. The statistical model is the brain. Everything else is signal.

1

Statistical Model — The Core Engine

A logistic regression trained on 22,762 MLB games (2015–2024) using gradient descent. Validated on a completely held-out 2025 season: 56.5% accuracy across 2,430 games with proper confidence calibration.

Elo Rating Gap

Team power differential

Pitcher FIP Diff

Fielding-independent pitching

Park Factor

Run environment adjustment

Game Context

Day/night, home advantage

2

Elo Ratings + Pitcher Power Index — Real-Time Power

Every team has a live Elo rating updated after each game, seeded from Pythagorean win expectation. Every starting pitcher has a Power Index (0–100) built from ERA, xERA, WHIP, K/9, quality start rate, and innings depth — with recency-weighted trend detection.

3

Vegas Consensus — Market Validation

The final prediction blends our statistical model (50%) with Vegas implied probability (50%). The market prices in information no model can capture — clubhouse dynamics, undisclosed injuries, late roster moves. We respect the line, then look for divergence.

4

AI Analysis Layer — Qualitative Edge

Claude AI reads 18 data dimensions per game — Statcast metrics, bullpen fatigue, weather, confirmed lineups, platoon matchups, head-to-head history — and produces the score prediction and reasoning. The AI doesn't pick the winner. The model does. The AI explains why.

Confidence Means Something

When we say 65%, we mean it. Calibrated on 2,430 holdout games.

Model Confidence Games Actual Win Rate Verdict
50–54%99453.5%Calibrated
55–59%83054.7%Calibrated
60–64%35756.9%Calibrated
65–69%19072.1%Underconfident ↑
70–74%5576.4%Underconfident ↑

Backtest: 2025 season holdout (2,430 games not used in training)

Self-Learning Pipeline

Diamond AI gets smarter every day. After each game is graded, the system recalculates Elo ratings, updates pitcher indexes, recalibrates confidence, and feeds corrections into the next prediction cycle. No manual intervention.

Hourly: Grade → Bankroll → Elo + PPI
Daily:  Auto-Learn → Edge Detection → Model Calibration
Game Day: Predictions at 8am, 10am, 12pm CT

Data Stack

Retrosheet

233K+ game logs (1871–2025)

Lahman DB

Season stats, team records, pitching

Statcast

xERA, exit velo, barrel rate, xwOBA

MLB Stats API

Live scores, lineups, rosters

Vegas Odds

Moneylines, totals, run lines

Claude AI

Qualitative analysis & reasoning