MLB V2 Is Live: Every Feature Powering Our Opening Day Model

⚾ OPENING DAY: MARCH 27

Three years ago, Donnie Dimes was a NCAAB model running picks on a spreadsheet. Today,
the teaser we dropped in March is a reality:
MLB V2 is fully built, backtested, and live for Opening Day.
This isn’t a beta. The crons are running, the picks are loading, and the O/U model is outputting
edges we’ve never had before.

Here’s a full breakdown of every feature powering the model when that first pitch crosses the plate Thursday.

🎯 Feature 1: Favorites-Only Moneyline

The biggest structural change in V2 is deceptively simple: we only bet favorites.

In our 2024 backtest, underdogs at the same edge threshold were consistently dragging returns down —
not because the model was wrong about the game, but because underdog payouts don’t justify the variance at
8% edge margins in baseball. Favorites at 8%+ were hitting at 60.9% with +5.1% ROI.
Underdogs at the same threshold? Dead weight.

So we cut them. Clean and simple. Every MLB moneyline pick from Donnie Dimes is a negative-ML favorite
with a meaningful model edge behind it. Volume is lower. Conviction is higher. That’s the trade we want.

📊 Feature 2: O/U Totals Market

This is the new market — and frankly, it might be the more exciting one.

The model now projects total runs for every game (home + away), compares that projection to the market’s
O/U line using a normal distribution, and calculates a win probability for both OVER and UNDER.
When the edge clears 6%, it becomes a pick. When it clears 10%, it becomes an ELITE pick.

The 2024 backtest on totals revealed something important: our UNDER signal is exceptional.
The model tends to project runs conservatively (because good pitching suppresses offense more than
offensive stats predict), which means our UNDER calls are systematically calibrated in a way that
OVERs aren’t — yet.

Every morning slate now includes both ML picks and O/U picks. You’ll see them labeled separately
on the daily card.

🐂 Feature 3: Bullpen ERA Layer

Starting pitchers get all the headlines. Bullpens win (or lose) games.

V2 incorporates bullpen ERA for all 30 teams as a late-inning suppression factor.
The logic: starters cover roughly 5.5 innings, bullpens cover the remaining 3.5.
A team with a 3.43 ERA bullpen (Dodgers) is systematically different from a team with a 5.12 ERA
bullpen (Rockies at altitude) — and that difference needs to be in the model, not ignored.

When both teams have elite bullpens, the model nudges the projected total down.
When both have weak bullpens, it nudges up. The adjustment is weighted by inning share
so it doesn’t overpower the starter signal — it refines it.

This is the same obsessive attention to layers that powers our nightly optimization process
on the NCAAB side — applied to baseball.

📈 Feature 4: Rolling In-Season Stats

This is the feature that makes V2 a long-season model, not just an Opening Day model.

For now, V2 is using 2025 season stats as the baseline — the most complete picture of each team
going into 2026. But as 2026 games are played, the model automatically starts incorporating
real in-season data. After 15 games per team, the rolling stats begin blending in.
By game 60 (roughly 37% through the season), the model has fully transitioned to live 2026 numbers.

What this means practically: the model gets sharper as the season progresses.
Teams that are over/underperforming their offseason projections get corrected automatically.
The model learns the 2026 season in real time. By June, the edges we’re finding will be based
on actual 2026 data — not preseason assumptions.

🏟️ Feature 5: Park Factor Adjustments (All 30 Parks)

Coors Field is not Target Field. Never was, never will be.

V2 applies park-specific run environment adjustments to every game projection.
Every stadium has a run factor (based on 3-year rolling FanGraphs data) that scales
projected scoring up or down based on the venue. Coors at 115% run factor.
Oracle Park in San Francisco at 95%. The full range across all 30 MLB parks.

This matters most for the O/U model: a game at Coors with a line of 9.5 means something
completely different than that same line at Petco Park. The park adjustment ensures the model
is measuring edges against the right baseline for each venue.

How Picks Are Delivered

Every morning at 8:00 AM PST, the model runs automatically.
It pulls that day’s lines from the market, evaluates every game, and generates a pick card with
moneyline picks and O/U picks — each labeled with tier, units, edge percentage, and projected total.

Every night at 10:00 PM PST, the nightly grade runs — scores backfill, picks get graded,
and a results card goes out to the channel with the day’s record and running season stats.

All MLB picks are delivered directly in the
Donnie Dimes AI Diamond Chat.
That’s where the NCAAB tournament picks are running right now, and where Opening Day MLB picks will post Thursday morning.

All picks are for research and entertainment purposes only. Past model performance does not guarantee future results.
Please bet responsibly and within your means. See our
full model methodology for more detail.