I've spent nearly a decade in big data at FAANG, working on some of the largest data sets on the planet. I've also loved baseball my entire life. This season I finally put the two together: machine learning, the cutting edge of modern sabermetrics, Statcast pitch-level data, Bayesian statistics, Monte Carlo simulation, and AI-generated analysis. The result is a single model that simulates every MLB game 10,000 times before first pitch and only bets when the math says the market is wrong.
One month in, here's what the receipts look like.
"The model isn't trying to pick winners. It's trying to find where the market is wrong. 30 days, 152 picks, and the gap shows up."
Run line picks
24–8 · 75%
+27.1% ROI · the model's strongest edge
Underdogs (+ money)
54–49 · 52%
+17.9% ROI · where the dollars come from
Best single day
+$6,920
Apr 16 · 5–1 on the board
Hottest stretch
+$26,022
Apr 8–19 · 12 days, 47–19 picks
⚙️
Monte Carlo simulation
Every MLB game simulated 10,000 times, plate appearance by plate appearance.
📊
Statcast + Bayesian regression
Pitch-level data (launch speed, launch angle, xwOBA, xBA), regressed for small samples and weighted by park and platoon splits.
💥
Hard-hit & barrel rates
Barrel %, hard-hit %, average exit velocity, handedness-split. We trust how hard the ball was hit, not BABIP luck.
🎯
Advanced K & whiff metrics
Strikeout rate, walk rate, whiff %, chase %, per batter and pitcher, vs L and vs R. The DNA of a plate appearance.
🧬
Cutting-edge sabermetrics
Skill-based player ratings from premier modern baseball research: contact quality, plate discipline, projected rate stats, layered into every plate appearance.
☁️
Weather & park factors
Wind speed, wind direction, temperature, and park-specific HR factors baked into every game's run environment.
🏃
Speed & baserunning
Sprint speed and baserunning ratings drive triples, doubles taken, and infield-hit probability. Small effects that stack.
🔄
Elo-blended priors
Team-strength signal merged with the simulation output so power-rankings disagreements get reconciled, not ignored.
🛡️
Advanced bankroll management
Fractional-Kelly stake sizing with edge thresholds, hard caps, and dynamic scaling. Smooth equity curve, no rocket-ship stakes.
"Quarter-Kelly. Five-percent cap. Smooth equity curve. Max drawdown was a single 15% day. The rest is just math compounding."
Phase 1: Mar 26 to Apr 7 (slow burn). Treaded water. +$752 in 13 days. The model wasn't broken; the variance just hadn't broken open yet. Discipline mattered more than picks.
Phase 2: Apr 8 to Apr 19 (the run). +$26,022 in 12 days. 47–19 on picks (71%). Five days over +$1,800. The model had been right all along; the slate finally cooperated.
Phase 3: Apr 20 to today (cooldown). Some give-back, but the bankroll is still nine consecutive sessions above $30K. Variance giveth, variance taketh; the process doesn't change.
All 152 picks of the season, every result, every dollar. Nothing edited after the fact. Hover any pick for the date, score, and P&L. Click any pick to see the full game analysis.