Pitch Intelligence
LAUNCH LAB
Loading…
Pitcher
pitches
pitch types
xwOBA against
deception score
Season
🔗
How does he sequence his pitches?
Flow width = how often that transition occurs. Hover any flow to see exact probability. Filter by count to reveal situational tendencies. ⚠ Season filter does not apply — Markov matrices use all-time data.
Count
Transition Detail
Hover a flow on the diagram to see details.
🚇
How deceptive is his arsenal?
A pitch pair tunnels well when both pitches look identical at the batter's decision point (~23 ft from plate) but arrive at different locations. Every combination in this pitcher's arsenal is scored below. Click any row to see the 3D trajectory visualization. Higher score = more deception.

Reading the 3D view: The Y-axis is distance remaining to the plate — pitches travel from back (~60ft) toward you (0ft). The ◆ gold diamonds mark the decision point at 23ft where the batter must commit to swing or take. Watch how close the two lines are at that point, then how they diverge toward the plate. Wide separation at the plate with tight similarity at the decision point = high tunnel score. Scores use velocity-adjusted movement similarity — a FF→SL pair scores differently than FF→CU even with similar geometry, because the speed gap is smaller and harder to detect by timing alone. Note on aggregation: each sequence score is computed from every real occurrence of that pitch pair in the data, then averaged. The pitcher deception score is a weighted average across all cross-type sequences, weighted by how often each sequence actually occurred.
Arsenal Tunnel Matrix
🚇
Select a Pitcher
Load a pitcher to score all their pitch pair combinations.
📐
Click a Row
Select a pitch pair from the matrix to visualize trajectories in 3D.
📋
What does his full arsenal look like?
Usage rate, velocity, whiff rate, and movement for every pitch. The sequences table shows which pitch-to-pitch combos produce the most or least damage. Green xwOBA = pitcher-favorable. Red = hitter-favorable. Best tunnel pair is highlighted at the bottom.
Select a Pitcher
Use the search bar above to load a full arsenal breakdown.
🏆
Who are the most deceptive pitchers in baseball?
Ranked by Deception Score — a weighted average tunnel score across all pitch pair combinations, weighted by how often each sequence actually occurs. Higher = harder to read out of the hand. Click any row to load that pitcher.
Filters
Min Pitches 150
Show
Top 20 — Deception Score
Loading…
Pitch Types
FF — 4-Seam FastballThe standard fastball. High velocity, minimal movement. The baseline pitch every other offering is designed to play off of.
SI — SinkerA fastball variant with arm-side run and downward movement. Generates ground balls. Often tunnels well with breaking balls.
FC — CutterA fastball with late glove-side cut. Harder than a slider but with less break. Extremely effective when tunneled off a 4-seam.
SL — SliderA breaking ball with glove-side lateral movement and some vertical drop. One of the most common strikeout pitches in MLB.
ST — SweeperA wide-breaking slider variant that sweeps horizontally across the zone. Became common post-2020. High whiff rates glove-side.
CU — CurveballPrimarily vertical drop with some lateral movement. Slower than sliders. Creates swing-and-miss when tunneled behind fastballs.
CH — ChangeupAn off-speed pitch designed to mimic fastball arm action but arrive 8–15 mph slower. Arm-side fade is common.
FS — SplitterFastball grip with fingers split wide, producing late downward dive. Common among power pitchers as a strikeout/chase pitch.
KC — Knuckle-CurveA curveball thrown with a knuckle grip, producing sharper late break. Smaller usage but high whiff potential.
Metrics
xwOBA (Expected Weighted On-Base Average)Measures the quality of contact on a per-pitch basis using exit velocity and launch angle, independent of fielding. Scale is roughly .250 (excellent pitcher) to .420 (poor). Lower is better for pitchers.
pfx_x — Horizontal MovementHow much the pitch moves horizontally relative to a theoretical spin-free pitch, measured in inches. Positive = arm-side, negative = glove-side (from pitcher's perspective).
pfx_z — Vertical Movement (IVB)Induced vertical break — how much the pitch rides up (positive) or drops down (negative) relative to gravity. High IVB on fastballs correlates with swing-and-miss above the zone.
Release SpeedVelocity measured at the release point, in mph. The Statcast measurement point is roughly 50 ft from the plate, slightly lower than perceived "gun" readings.
Whiff RateSwings-and-misses divided by total swings. A high whiff rate (>30%) on any pitch indicates a strong swing-and-miss offering. League average is roughly 25–27%.
Usage %How often a pitcher throws each pitch type, as a percentage of total pitches. Reflects pitch mix tendencies across all counts and situations.
Tunneling Concepts
TunnelingThe principle that two pitches are most deceptive when they travel along the same path (the "tunnel") for as long as possible before diverging. A batter must commit to swinging before the divergence point.
Decision Point (~23 ft)The approximate distance from the plate at which a batter must commit to a swing or take. Pitches that look identical at this point — regardless of where they end up — are the hardest to distinguish.
Tunnel DistanceThe spatial separation between two pitches at the decision point (23 ft from plate), measured in feet. Lower = better tunnel. Ideally under 0.3 ft for elite pitch pairs.
Plate SeparationThe distance between where two pitches cross the plate, in feet. Higher plate separation combined with low tunnel distance = high deception.
Diverge RatioPlate separation divided by tunnel distance. How much the pitches "spread apart" after the decision point relative to how similar they looked before it. The core driver of tunnel score.
Tunnel ScoreA 0–100 score for a specific cross-pitch-type sequence. Combines plate separation, movement vector similarity at the decision point, release point similarity, and a velocity similarity penalty — pitches with a large speed gap (e.g. FF→CU at 15mph diff) score lower because batters can identify them by timing alone. Scores above 40 represent strong tunneling; 20–40 is moderate; below 20 is low deception.
Deception ScoreA pitcher's aggregate tunnel score across their entire arsenal. Weighted by how often each pitch pair actually appears in real sequences — so a frequently-thrown FF→SL combination contributes more than a rare pairing. Scale: 0–100.
Models & FAQ
How is the xwOBA model built?An XGBoost regressor trained on Statcast pitch-level data. Features include release speed, velocity change from previous pitch, vertical break difference, count state, handedness-adjusted movement, and at-bat depth. Trained with time-series cross-validation to prevent data leakage.
How do the Markov matrices work?For each count state (0-0, 0-2, 3-2, etc.), a transition matrix is built from real pitch sequences in the data. Each cell represents the probability of pitch B following pitch A in that count. Count-conditional matrices are used when available; a global matrix is the fallback.
Why are Markov probs filtered to the pitcher's arsenal?The matrices are built on league-wide data. When viewing a specific pitcher, probabilities are filtered to their actual pitch types — so a pitcher who doesn't throw curveballs won't show curveball probability.
What does "EV Mode" mean on the Next Pitch tab?Expected Value mode. Enter the pitch that was actually thrown to see how predictable it was — specifically, where it ranked in the Markov probability distribution. Rank #1 means it was the most likely pitch given the count and previous pitch.
Why does the Season filter not affect Markov or xwOBA?The prediction model and Markov matrices benefit from as much historical data as possible. They are trained on all available seasons. The season filter only affects pitcher-level profiles — pitch mix, velocity, movement, and deception scores — where recency matters more.
How is Starter vs Reliever determined?A heuristic based on pitch arsenal size and volume: pitchers with 4+ distinct pitch types or 400+ pitches in the dataset are classified as starters. All others are relievers. This is approximate — edge cases exist.
What data source is used?Baseball Savant (MLB's Statcast portal) via direct CSV download. Data includes pitch-level tracking for all MLB regular season and postseason games. Updated by re-running the pipeline.
🔬
What does the data actually tell us?
Twelve analyses computed live from the loaded dataset using pitch-level tunnel scoring with velocity similarity weighting. Key finding: movement profile similarity drives deception — pitchers whose pitches move and travel similarly are hardest to identify at the decision point. All findings update with the Season selector. Min pitches: 150.
Loading…