RoboNiner is a multi-layer career projection system covering MLB players and all levels of the minor leagues. It generates Y+0 through Y+18 (or until career-end) projections for every hitter and pitcher in the affiliated minor league system.

Design Goals

Understanding the major public systems clarifies RoboNiner's design choices.

PECOTA (Baseball Prospectus)

Method: Comparables-based. Finds 25 historical players most similar to the target via Mahalanobis distance (age, level, skill rates, body type, handedness), then blends their actual career trajectories. Produces percentile bands (10th/90th).

ZiPS (Dan Szymborski, FanGraphs)

Method: Marcel-based (3-year weighted average with regression), modified by Tango/Lichtman aging research. Comps inform aging curves rather than trajectories directly.

Steamer (Jared Cross, FanGraphs)

Method: Component-based regression with aging. Separately models AVG, OBP, SLG, K%, BB%, BABIP using independent regression weights. Consistently the most accurate single-year system in public evaluations.

THE BAT X (Derek Carty)

Method: Neural network / ML approach on Marcel foundation. Most aggressive Statcast integration of any public system.

OOPSY (Dave Fleming)

Method: Similar to ZiPS with validated young-for-level adjustments. Key insight: a player 2+ years younger than the average at their level has demonstrated elite talent relative to peers.

Projections are built sequentially. Each layer adds information the prior layer lacks.

Layer 1: Marcel+ Foundation

Base: Tom Tango's Marcel system (5/4/3 year weights) extended with stat-specific regression. K% and BB% regress quickly (high year-to-year correlation); BABIP regresses heavily (noisy, environment-sensitive):

Young-for-level bonus (from OOPSY research): a player 2+ years younger than the league average at their level has demonstrated elite talent relative to peers. Regression shrinkage is reduced up to 80% for players 3+ years young for their level:

ageRelativeBonus = 1.0 + min(0.80, yearsYoung × 0.28)
averageAges: ROK=19.0, A=21.5, A+=22.5, AA=23.5, AAA=26.0, MLB=28.5

Minor league regression multiplier: minor league stats are noisier than MLB. Regression is amplified by level:

Layer 2: Component Rate Decomposition

Rather than aging AVG/OBP/SLG directly, we decompose to rate components and age each independently. Hitters: K_rate, BB_rate, HR_rate, BABIP, 2B_rate, 3B_rate, SB_rate, HBP_rate. Pitchers: K_rate, BB_rate, HR_rate, H_rate, LOB%.

MLEs convert minor league stats to their MLB equivalent. A .310 AVG in AAA is worth roughly .297 in MLB. A .310 AVG in Low-A is worth roughly .240.

How MLEs Are Computed

For each year/level, we collect players who appeared at that level and MLB within the same season (called-up players), then compute the ratio of their rate stats at each level.

⚠ The Selection Bias Problem

Population-Level Caps

We apply maximum multiplier caps reflecting what the average player at each level would translate to:

Park factors are applied at the component level, not just summary stats. Coors Field boosts HR differently from BABIP.

Computation Method

MLB + AAA: PF = (home_stat / home_PA) / (away_stat / away_PA) × 100 using homeAndAway splits from the MLB Stats API. Three-year weighted smoothing (5/4/3) reduces season noise.

AA/A+/A/ROK: Single-regression approximation: PF = (team_stat_rate / league_avg_rate) × 100. Less accurate but adequate at the aggregate level.

Application: parkAdjusted = observed / (1 + (PF - 100) / 200). The /200 denominator (not /100) reflects that players play ~50% of games at home.

Notable Parks (Current Season)

Beyond overall level factors, we compute per-league adjustments within each level. The PCL (Pacific Coast League, AAA) is historically hitter-friendly; the IL (International League) runs closer to MLB environments.

Step 1: Find Historical Comparables

For each projection target, find 25 players from the Lahman database (1955–2025) most similar at the target's current age. Similarity measured via Mahalanobis distance:

Step 2: Extract Comp Aging Curves

For each comp, trace year-over-year stat changes weighted by: sample size (PA/BF) and comp distance (1 / (0.5 + distance)). Result: per-stat, per-age aging deltas for K_rate, BB_rate, BABIP, HR_rate, ISO, SB_rate.

Step 3: Blend Comp vs. Baseline

if compConfidence >= 0.5:
  delta = comp_delta × blend + baseline_delta × (1 - blend)
  blend = min(0.80, compConfidence)
else:
  delta = baseline_delta  # fall back to research-backed fixed curves

Career endpoints from attrition: At each age, we track what fraction of the comp pool was still active. When fewer than 20% of comps are still playing AND OPS < .550 (or ERA > 6.50), the player retires.

Baseline Aging Parameters (Fallback)

Sources: Tango/Lichtman (The Book, 2006), FanGraphs aging studies, Lichtman's OOPSY research.

Position-Specific Aging Multipliers

Multipliers ramp from 1.0 at age 22 to full value at 27 (development phase shouldn't apply a wear penalty).

Velocity Aging (Pitchers)

Pitchers begin losing ~0.25 mph/yr of fastball velocity after age 25. Each mph lost reduces K_rate:

veloLost   = (age - max(startAge, 25)) × 0.25  // mph
kRateLoss  = veloLost × 0.0008                  // K/BF reduction per mph

This compounds with comp-driven curves: the curves capture the empirical K decline; velocity aging explains the mechanism.

Statcast calibrates the Marcel+ foundation with objective physical measurements more predictive than rate stats alone.

Hitter Inputs

Pitcher Inputs

OAA (Outs Above Average)

Multi-year Marcel-weighted OAA from Baseball Savant (3-year, 5/4/3 weights), regressed toward 0 using position-specific reliability:

Catcher Framing

Framing runs from Baseball Savant runs_extra_strikes, regressed 50% toward 0. Age curve: improving through 28, peak 28–30, steep decline after 30 (~1.2 runs/yr).

Position Transitions

As players age, premium defenders move to easier positions. Interpolated over the transition window, not as a cliff:

Sprint Speed Modifier

speedRetention = max(0.70, min(1.15, 0.70 + (sprint_speed / 27.0) × 0.30))

Elite speedsters (30+ ft/s) retain defensive value longer. Slow players (24 ft/s) lose range value faster.

Hitter PA Arc

Comp PT blending (when ≥5 comps available):

blendedPA = formulaPA × 0.40 + compAvgPA × 0.60
blendedPA × compAttrition.pctActive  // adjusts for career-end probability

Career Endpoints

Career ends when:

1. Fewer than 20% of comparables still active at that age AND OPS < .550 (hitters) or ERA > 6.50 (pitchers)

2. Hard ceiling: age 48

WAR = wRAA + BsR + Fld + Frm + Pos + Rep

// wOBA (FanGraphs linear weights)
wOBA = (BB×0.690 + HBP×0.720 + 1B×0.880 + 2B×1.245 + 3B×1.575 + HR×2.015) / PA

// wRAA (value above average)
wRAA = (wOBA - lgwOBA) / wOBAscale × PA

// BsR (baserunning)
netSB = SB - (CS × 2)
wSB   = netSB × 0.2    // ~0.2 runs per net steal

// Replacement level
hitter: 20.0 replacement runs per 600 PA
pitcher: 18.0 replacement runs per 200 IP

// Runs per win
RPW = 10.0  // FanGraphs scale

High-Priority Future Work