MLB V3.5: We Taught Our Baseball Model to Think in Over/Unders

If you’ve been following Donnie Dimes, you know we don’t ship model upgrades without a reason. V3.5 is the biggest improvement to our MLB engine since launch — and it started with a simple question: why are we projecting run totals without knowing whether these teams actually hit their numbers?

The Gap We Identified

Our V2 model was already solid. It projected run scoring using rolling team stats, starting pitcher FIP vs team ERA, bullpen adjustments, park factors, and home field advantage. On Opening Day, it correctly identified three OVER plays — including a +2.21 run edge on the Dodgers/D-backs game and an 11-run Cardinals/Rays total that hit comfortably.

But it was missing something. Two teams can both average 4.5 runs per game, yet one team consistently plays 9-run slugfests while the other grinds out 5-2 final scores. The market prices this in. Our model wasn’t.

What V3.5 Adds

We built a new module — rolling_ou_stats.py — that tracks two new signals and feeds them directly into every totals projection:

1. Team O/U Hit Rate (L10 + L5)

Every team now carries an over/under hit rate for their last 10 games and their last 5 games, plus a scoring delta that measures how much their recent run output deviates from their season average. This gets applied as a small multiplier to the projected total — conservative by design, capped at ±10%, but directionally meaningful.

2. Pitcher O/U Tendency (L8 Starts)

Some starters are run-suppression machines. Others are fireworks shows waiting to happen. V3.5 now tracks each pitcher’s last 8 starts and factors that O/U history into the total projection. As 2026 data accumulates game by game, this signal sharpens automatically.

How It Works in Practice

Take the Cleveland @ Seattle projection from March 29th:

The V3.5 model pulled back the projection by 10% because both Seattle and Cleveland had been going UNDER heavily in recent games — SEA at 40% OVER rate in their last 10, CLE at 30%. The model correctly noted that and adjusted. The game still hit the OVER. Both versions got it right, but V3.5 gave a more honest number.

Early O/U Trends Worth Watching

The model is already surfacing meaningful patterns from 2025 historical data — and these update daily as 2026 games roll in:

Ghost Mode — Still Earning It

MLB picks are still internal only. We don’t post publicly until the model demonstrates a positive return on units with real picks over a meaningful sample. That’s not a hedge — that’s how we operate. Our NHL model went through the same process before going live.

V3.5 is the foundation we wanted before flipping the public switch. The data collection is live, the O/U signals are running daily, and the track record is building. When we go public, it’ll be because the numbers earned it.

Follow the Donnie Dimes AI Diamond Chat to be the first to know when MLB goes public. 🔒

Frequently Asked Questions

What is MLB V3.5?

MLB V3.5 is an upgrade to the Donnie Dimes baseball prediction model that adds team over/under hit rate tracking and pitcher O/U tendency analysis. The model now considers how often each team’s games go OVER or UNDER in their last 10 games, plus each starting pitcher’s historical tendency across their last 8 starts.

How does the team O/U hit rate work in the model?

The model tracks each team’s OVER and UNDER hit rate over their last 10 and last 5 games, plus a scoring delta measuring how much recent run output deviates from the season average. This is applied as a multiplier (capped at ±10%) on every projected total.

When will Donnie Dimes MLB picks go public?

When the model shows consistent positive units on real picks over a meaningful sample. We’re collecting live 2026 data in ghost mode right now. Follow the Diamond Chat to be the first to know.