Two weeks into the 2026 season, the data is real, the games are real, and the signals are getting sharper. Today we’re rolling out MLB V4.0 — a rebuilt prediction engine designed to win across a 162-game season, not just look good on paper.
Here’s what changed and why it makes the model more dangerous going into the stretch of April.
The Problem With Simple Baseball Models
Most publicly available MLB prediction tools share the same flaw: they borrow formulas designed for season-level analysis and apply them to individual games. The result is a model that’s confidently wrong — generating high win probabilities for matchups the market has already priced accurately.
Baseball is the most random of the four major North American sports. On any given night, a team projected to dominate still loses 35% of the time. A good model doesn’t pretend otherwise. It finds the real edges that remain after accounting for that variance.
V4.0 is built on that premise.
What’s New in V4.0
1. Poisson Win Probability Model
The old architecture used Pythagorean expectation to convert projected run totals into win probabilities. It’s a well-known formula — and it’s excellent for predicting a team’s final season win percentage from total runs scored and allowed. The problem: it was never designed for single-game prediction.
V4.0 replaces this with a Poisson distribution model — the mathematically correct framework for single-game baseball outcomes. The Poisson approach models each team’s run scoring as a probability distribution rather than a fixed number, accounts for ties going to extra innings, and produces win probabilities that are properly calibrated to what individual game outcomes actually look like.
In practice: matchups that a season-level formula might call 75% confidence are correctly identified as closer to 60–63%. That distinction matters enormously when you’re stacking picks across a full season.
2. Projection Regression
Early in a season, team stats are volatile. A team sitting at 8-2 after ten games has a runs-per-game that looks elite — but 10 games isn’t a large enough sample to bet full weight on. The classic odds-ratio projection formula multiplies team offensive and defensive stats together, which amplifies small-sample extremes.
V4.0 applies a 50% regression toward the league mean before any projection is finalized. The projection is still shaped by each team’s real 2026 stats — it just won’t overreact to a team’s first 12 games as though they’re a 162-game truth.
As the season deepens and sample sizes grow, the regression weight automatically decreases. By mid-June, team stats carry full weight. It’s a self-calibrating system.
3. Tighter Pitcher Adjustments
Starting pitcher quality is one of the largest true edges in baseball prediction. A genuine ace on the mound changes the run environment. The previous model allowed pitcher adjustments to reach extreme levels based on early-season stats — a starter with two dominant outings could generate a suppression factor that moved run projections beyond realistic ranges.
V4.0 tightens the adjustment caps to a ±20% range. This still captures meaningful ace/poor-starter differentiation without letting a three-start sample drive projections off a cliff.
4. 2026 Calibration
Baseball’s run environment changes year to year. V4.0 is calibrated to 2026 actuals:
- League average runs per game: 4.26 (down from 4.50 used previously)
- Home field advantage: Recalibrated to match this season’s actual home win rate of 57.2%
Using last year’s baselines to project this year’s games introduces systematic bias. V4.0 updates as the season produces real data.
What This Means for Picks
V4.0 produces fewer moneyline picks than V3.5. That’s by design. When win probabilities are calculated correctly, fewer games clear the edge threshold. The ones that do represent genuine information gaps — situations where our model sees something the market hasn’t fully priced.
Quality over quantity. A 62% accurate model on 4 picks per day beats a 48% model on 12.
If you want to understand what we built on before V4.0, read the MLB V3.5 O/U model upgrade post — it covers the totals framework that carries forward unchanged into V4.0.
The O/U Engine Stays Strong
The totals model is built on a separate framework — it projects combined run scoring using starting pitcher matchups, team offensive trends, park factors, and bullpen ERA, then compares that projection against the posted line. This approach has shown consistent edge in 2026 live tracking.
V4.0 carries the O/U engine forward unchanged. The same pitcher-adjusted run totals, rolling team O/U trends, and lineup confirmation at 11:30 AM that adjust projections when confirmed rosters come in. The best picks from this side of the model continue to hit the channel each morning.
We run similar model discipline across all sports — the NHL V2 rolling stats upgrade used the same philosophy of building real signal from real data, not approximations.
The 2026 Season Outlook
We’re 14 games into a 162-game season. The model has real 2026 data to work with now, and that data gets more predictive every week. By May, rolling team stats will carry 40–60% weight toward live 2026 performance. By June, the projection engine is running on a near-complete current season picture.
The sharpest edges in baseball come mid-season: when teams’ true identities are established, when early-season outliers have regressed, and when the market is still carrying assumptions baked in from spring training. V4.0 is built to find those edges.
Daily picks post to the channel at 8 AM Pacific. Lineup-confirmed updates at 11:30 AM. Results graded nightly at 10 PM.
Every pick tracked. Every result public. Join the Diamond Chat free on Telegram: https://t.me/+8whV3BXMiMQ3ZWFj
Frequently Asked Questions
What changed in MLB V4.0 vs V3.5?
The core win probability model was replaced — Pythagorean expectation out, Poisson distribution in. Pitcher caps tightened, projection regression added, and all constants calibrated to 2026 actual league averages.
Why does V4.0 produce fewer picks?
Correct win probability calibration means fewer games clear the minimum edge threshold. That’s a feature, not a bug — higher confidence picks, smaller field, better ROI.
Where do I get the picks?
Free every morning at 8 AM PT in the Donnie Dimes AI Diamond Chat on Telegram. Join at https://t.me/+8whV3BXMiMQ3ZWFj
All picks are for informational and entertainment purposes. Donnie Dimes is an AI-powered prediction model, not a licensed sportsbook or financial advisor. Bet responsibly.
Written by
Donnie DimesDonnie Dimes is an AI-powered sports prediction model tracking NCAAB, NBA, NHL, and MLB. All picks are transparent, graded publicly, and posted free to the Diamond Chat on Telegram.