NRFI / YRFI methodology
How the NRFI / YRFI model works
Probability of zero combined runs in the first inning, scored from each starter's 1st-inning history plus park, ump, and weather context.
Inputs
- • Starter 1st-inning rate stats (NRFI%, ERA, K%, WHIP, FB velo, whiff rate)
- • Park run / HR / hit factors
- • Umpire zone size + K factor + NRFI rate (from /mlb/umpires/)
- • Weather (temp, wind, humidity, roof state)
- • Lineup status (projected vs confirmed)
Outputs
- • Probability of NRFI per game (0..1)
- • Lean label (Strong NRFI / Lean NRFI / Neutral / Lean YRFI / Strong YRFI)
- • Whether the read is strong enough for the public picks board or better left as research context
- • When available: the sportsbook line and price the model compared against
Approach
NRFI scores each game by combining both starting pitchers' first-inning rate stats with park, umpire, and weather adjustments. The output is a probability that no run scores in the top or bottom of the first.
YRFI is the complement plus an asymmetric adjustment for high-volatility starts (low-information rookies, late lineup uncertainty). The two outputs share the same core feature set — they differ in how they treat tail risk.
Training data
Multi-year historical first-inning outcomes from MLB StatsAPI box scores, stitched to per-pitcher 1st-inning stats and Statcast metrics.
Holdout: most recent ~10% of games are reserved as a temporal holdout each retrain. We do not shuffle across seasons.
Calibration
After games are graded, the model is checked against what actually happened and adjusted so a 60% number behaves like a real 60% over time.
NRFI has the longest public track record on the site, so its probabilities are the most mature of the current model group.
How a pick becomes official
Official picks need a complete pregame record: the matchup, the model side, the available sportsbook price, and the line the model was comparing against.
If that record is incomplete, the pick can still be useful context, but it stays in research instead of counting toward the public record.
Known biases and limits
- ! Early in the season, individual starter samples are small. The model widens uncertainty bands and is more cautious about what counts toward the public record.
- ! NRFI is the most-sampled model and therefore the most trustworthy public record. YRFI is much less frequently the top pick and should be treated as supplementary signal.
Not for use as
- × Predicting one exact at-bat or one exact inning. These are probability models, not play-by-play fortune tellers.
- × Ignoring the sportsbook price. A model edge only matters if the available price is good enough.
- × Replacing late-breaking baseball judgment. Lineups, weather, scratches, and pitching changes can move after the page refreshes.