MLB Prediction Models Explained: How Projection Systems Inform Betting Decisions

I built my first MLB prediction model on a spreadsheet in 2019. It was terrible — a crude formula weighting pitcher ERA, team batting average, and home-field advantage into a single number that I compared against the moneyline. The model lost money for two months before I shelved it and started studying how professionals actually build projection systems. What I discovered was that the publicly available models are far more sophisticated than anything I could build alone, and the real skill is not in building a model but in knowing how to use one — and when to override it.

Publicly Available MLB Projection Systems: What They Cover

Several well-regarded MLB projection systems are available for free or at low cost. They use different methodologies, weigh different variables, and produce different outputs, but they share a common goal: projecting the expected performance of players and teams more accurately than raw statistics can.

FIP is a better predictor of future pitching performance than ERA, and the best projection systems take this principle and extend it across every dimension of the game. They project pitcher performance using multiple seasons of data, regression toward the mean, ageing curves, and underlying component stats. They project hitter performance using platoon splits, park-adjusted metrics, and batted-ball quality. They combine these player-level projections into team-level win probabilities for each game.

The most widely referenced public system in the betting community is one that generates win probabilities for every game based on projected starting lineups, starting pitchers, and park factors. These probabilities are published daily and updated as lineup information becomes available. Other systems focus specifically on player projections, which bettors can then apply to prop markets or aggregate into their own game-level models.

What these systems do not do is account for the betting line. They project outcomes; they do not assess value. A model might project Team A with a 58% win probability, but if the bookmaker is pricing Team A at an implied 60% probability, the model says Team A is overpriced despite being the “right” side. That distinction between the model’s projection and the bookmaker’s implied probability is where betting value is determined — and it is the step that requires human judgement.

Expected Value Methodology: How Models Identify Mispriced Lines

Prediction markets have diverted more than $500 million in potential tax revenue from traditional sportsbooks in the past year, partly because alternative platforms offer a different way to express probabilistic views. The rise of prediction markets underscores a broader truth: the sports betting ecosystem is built on probability estimation, and the most successful bettors are those who estimate probability more accurately than the closing line.

Expected value — EV — is the mathematical framework that connects model projections to profitable betting. EV asks: if I made this bet a thousand times at these odds, would I profit or lose? The formula is: EV = (win probability x profit if win) – (loss probability x stake). A positive EV means the bet is profitable over the long run. A negative EV means it is not.

To calculate EV on an MLB bet, you need two inputs: your estimated win probability and the bookmaker’s odds. If your model projects Team A at 55% and the bookmaker offers 2.00 (implying 50%), the EV is: (0.55 x 1.00) – (0.45 x 1.00) = +0.10, or +10% of your stake. That is a strong positive-EV bet. If the bookmaker offers 1.75 (implying 57.1%), the EV is: (0.55 x 0.75) – (0.45 x 1.00) = -0.0375, or -3.75%. That is a negative-EV bet despite your model saying Team A is the more likely winner.

This distinction is the most important concept in model-driven betting. You are not betting on who will win. You are betting on whether the price is right. A bet on a 35% underdog at 4.00 (implying 25%) is a better bet than a wager on a 60% favourite at 1.55 (implying 64.5%). The first has positive EV; the second has negative EV. Models give you the probability; the price determines the value.

Limitations of Models: What Algorithms Cannot See

Andrew Rhodes, the UK Gambling Commission’s chief executive, has talked about discussions with operators showing a widening of the sports offering in the UK, with sports beyond football and horse racing growing in popularity. As MLB becomes more visible to UK bettors, the temptation to rely entirely on models — outsourcing the thinking to a projection system — grows. That temptation should be resisted.

Models cannot see clubhouse dynamics. A team in the middle of a managerial firing, a locker room divided by a contract dispute, or a squad galvanised by a teammate’s personal tragedy — these human factors influence performance but are invisible to algorithms. I have overridden my model’s projection on multiple occasions based on situational awareness that the model could not capture, and those overrides have been among my most profitable bets.

Models also struggle with regime changes. A new pitching coach who revamps a pitcher’s approach, a new hitting philosophy that changes a lineup’s aggressiveness, a rule change like the pitch clock that reshapes game dynamics — these shifts take time to appear in the data, and models that rely on historical data will lag the new reality. Early in a season, or immediately after a major coaching change, models are at their weakest.

Weather, travel, and scheduling quirks are partially captured by some models but often imprecisely. A model might adjust for park factors but not for the specific wind conditions on a Tuesday afternoon. It might account for home-field advantage but not for a team playing its sixth game in six days after a cross-country travel day. These gaps are where human analysis adds value on top of model output.

Using Models as One Input, Not as the Entire System

My current approach uses a public projection system as my starting point. The model’s win probability gives me a baseline, which I then adjust based on factors the model misses: bullpen fatigue, lineup changes, weather, umpire assignment, and any situational factors I have identified through my own research. The final probability I use for calculating EV is rarely identical to the model’s output — it is my synthesis of the model’s projection and my own analysis.

When my adjusted probability and the model’s projection agree, I bet with higher confidence. When they disagree significantly, I investigate the source of the disagreement. If the model sees something I missed — a platoon disadvantage I overlooked, a park factor I underweighted — I adjust toward the model. If I see something the model missed — a bullpen running on empty, a pitcher who looked injured in his last start — I adjust toward my own assessment.

The worst outcome is blindly following a model without understanding its inputs or limitations. A model is a tool, and like any tool, its value depends on the skill of the person using it. The bettor who understands why the model projects what it does, and can identify the situations where the model is likely to be wrong, will outperform the bettor who simply bets every positive-EV output without critical assessment. A complete MLB betting framework integrates model output with matchup analysis, market reading, and bankroll management into a coherent system — the model is one pillar, not the entire structure.

Published by the DiamondEdge team.