How Databaseball Works

Databaseball runs two independent predictive models trained on multiple years of MLB pitch-level data:

• Starter Ks: predicts each starting pitcher's total strikeouts and computes over/under probabilities at standard K lines (3.5 through 8.5). This is the primary model.
• First Five Innings (F5): predicts the run differential after five innings and produces home/away/tie win probabilities.

Each model follows the same general pipeline: raw pitch-by-pitch data is transformed into statistical profiles for pitchers, opposing lineups, and game context. Environmental factors including weather conditions and home plate umpire tendencies are incorporated as contextual signals. A machine learning model produces an expected value, which is then converted into probabilities.

The two models differ in how they produce probabilities from the predicted expected value. The Ks model uses an analytical probability distribution to compute P(over) at any line directly. The F5 model uses 10,000-draw Monte Carlo simulation to produce three-way win probabilities (home, away, tie), since the interaction between two predicted run totals requires simulation rather than a closed-form solution.

When confirmed lineups are posted (typically 1–2 hours before game time), predictions are automatically recomputed with lineup-specific data and an updated weather forecast closer to game time. The card updates to show "Lineups ✓".

A prediction is only useful if the probabilities mean what they say. Databaseball validates this by bucketing all historical predictions by model confidence and measuring actual accuracy.

The calibration sidebar on each model page shows these results directly. For the Ks model at the 5.5 K line (the most common prop):

• Picks at ≥55% confidence hit at 69.6% accuracy
• Picks at ≥60% confidence hit at 72.0% accuracy
• Picks at ≥65% confidence hit at 74.5% accuracy
• Picks at ≥70% confidence hit at 77.4% accuracy

These numbers come from 3,907 starter-games in the 2025 season. Early 2026 live results have tracked within 1–3 percentage points of the backtest.

Calibration measures whether the model's probabilities are accurate. Closing line value (CLV) measures whether the model knows something the market doesn't.

Across 7,264 matched predictions in the 2025 season (verified against closing odds from 9 sportsbooks), the Ks model carries an average CLV of +3.1 percentage points over the consensus no-vig closing line. 63% of picks have positive CLV.

CLV scales monotonically with model confidence:

• 50–55% confidence: ~0pp CLV
• 65–70% confidence: +2.5pp CLV
• 75–80% confidence: +5.7pp CLV
• 80%+ confidence: +9.3pp CLV

This means the model's predictions carry real informational edge: higher confidence picks correspond to larger market mispricings, not just directional accuracy.

The model agrees with the market's favored side 85.7% of the time. Its edge comes from refining the market's view with more precision, not from contradicting it.

Predictions are regenerated every 2 minutes during game hours (11am–1am ET) and every 30 minutes outside that window. Feature data is updated daily from Statcast.

Databaseball is a research and analytics tool. The probabilities represent the model's best estimate given available data, not guarantees. The calibration data is provided so users can evaluate the model's track record and decide for themselves what confidence level warrants attention.