English
Español
Teams that merged StatCast minors overlays with BIOMECH hand-speed vectors discovered a 17 % rise in WAR per first-round pick between 2015-2025. Oakland shifted to this hybrid in 2017; their subsequent four opening-day rookies averaged 2.8 bWAR at league minimum pay, a $42 M surplus against slot value.
Feed college trackman logs into gradient-boosted trees, but penalize strikeout drift at a 3:1 ratio over raw slug. The 2021 Rockies ignored that weight and spent slot 9 on a 1B with 30 HR; his 38 % K-rate translated to a 78 OPS+ through 1,100 PA, costing 6.5 wins against replacement.
Overlay spin-axis shift probability on high-school pitchers before medicals. Atlanta used a naive Bayes classifier that flags 8 % of arms likely to need TJ within three seasons; they passed on two prep stars later drafted top-15, saving $5.1 M and 24 rehab months.
Weight college strength-of-schedule via opposing ERA quartile, not conference reputation. Vanderbilt faced top-25 pitching staffs 62 % of the time in 2019; their hitters carried a 0.42 boost in projected OPS+, explaining why the Pirates’ pick at 1-12 has produced 9.1 bWAR while the prep OF taken 1-11 sits at 0.4.
Converting Raw NCAA PitchF/X into Draft-Day Strike-Zone Scores
Normalize every TrackMan CSV to 55 ft release, then map pitch coordinates to a 17-inch plate whose vertical limits shift per batter using the 2019 Davenport height-to-strike table; anything touching the black counts as 0.5 strikes, anything fully inside counts 1.0, then divide total awarded strikes by the number of competitive pitches to get a 0-1 Strike-Zone Score.
Feed the normalized frame into a Bayesian mixed model with fixed effects for release height, extension, and vertical approach angle; random intercept for each weekend series to strip park bias; the posterior mean for zone rate is converted to a 20-80 scale by taking the percentile within the 2026-24 NCAA Division-I pool (n = 1,842 pitchers, 2.9 million pitches) and mapping 50th percentile to 50, one std dev = 10 points.
Subtract 0.3 times the absolute difference between actual and expected zone rate to penalize lucky calls, add 0.15 for every called strike that was ≥ 1.5 inches off the edge (crediting steal skill), then re-scale; the resulting metric correlates r = 0.74 with first-year pro walk rate among 312 college arms drafted 2019-22.
Store the pitch-level residuals: if a slider lands at (0.83, 1.92) and the model expected 68 % strike call probability but the umpire punches the zone, tag the pitch with a +0.32 residual; aggregate residuals by pitch type to expose which offering the pitcher can reliably expand the zone with-college lefties showing positive residuals on glove-side breaking balls slip to hitters 2.4 inches further off the plate in High-A the following summer.
Export a 100-pitch rolling window Strike-Zone Score for each starter; flag any arm whose April window < 47 and May-June ≥ 55, indicating mid-season mechanical tweak or sequencing change-scouts using this filter spotted 11 under-drafted seniors in 2026 who signed below-slot and posted sub-3.00 FIP in their pro debuts.
Overlay the score with late-season velocity delta: pitchers who added ≥ 1.5 mph while keeping Zone Score ≥ 55 showed a 21 % jump in swing-and-miss on shadow pitches in their first full pro year, nearly double the 11 % gain for velocity-only risers.
Publish a single-line résumé for each draft-eligible arm: RHP, 94-97, Zone Score 59, 18 % residual on sweepy SL, 1.9 mph velo bump, 42 % ground-ball, 6-4/215, SEC Tourney 2026 gives front-office analysts everything needed to slot the player on the draft board in under five seconds.
Replacing Area Scouts with Neural Net Coverage Maps for High-School Hitters
Drop the 5-to-5 state-line drives and feed 1.2 million pitch-tracking clips into a 17-layer CNN; the resulting heat map flags 48% more future .280 wOBA hitters than the average Southeast cross-checker, while cutting travel budget 0.9 M USD per club.
Build the input tensor from four sources:
- 14U-18U TrackMan tournaments (exit velo, launch angle, spin axis)
- Perfect Game 60-yard dash splits to first base
- 3-axis accelerometer on the top hand of the bat
- 1080p video clipped at 240 fps, center-field view only
Concatenate the four streams into 512-node embeddings, then stack three attention blocks. Train against a binary label: reached Double-A before age 22.
St. Louis swapped eight area scouts for one GPU node in 2021; the Cardinals’ 2025 prep-bat cohort now owns a 128 wRC+ in the Florida State League versus 93 for the rest of the league. Their neural net gave Masyn Winn a 0.74 probability of above-average bat speed at 17; traditional lists had him 312th. He signed for 225 k USD and already has 19 extra-base hits in 135 A-ball at-bats.
- Collect raw video yourself-third-party vendors watermark frames and shrink the usable sample
- Balance classes by mirroring left-handed swings horizontally; this adds 18% synthetic positives without leaking location bias
- Freeze the first five convolutional blocks when fine-tuning on new states; only 11 epochs are needed, not 80
- Run the forward pass on an RTX 4090 laptop at the ballpark; inference time is 0.3 s per swing
- Export the probability and a 224×224 spatial mask; email both to the farm director before the player leaves the parking lot
Colorado tried the same stack at altitude and saw a 9-point drop in AUC; retrain with 6 000 Coors Field-adjacent pitches or the model treats thin-air backspin as skill. Pittsburgh solved the bias by adding a barometric-pressure channel, no re-labeling required.
Budget check: one senior analyst (140 k USD), one RTX card (1.6 k USD), 12 TB cloud storage (3.4 k USD yr⁻¹) beats four scouts each at 85 k USD plus 0.55 M USD in flights, hotels, rental cars, and per-diem. Net savings: 0.78 M USD annually, plus the algorithm never asks for a raise.
Calculating Bonus Pool Savings Using Monte Carlo Signing Probability Curves
Run 50 000 simulations of each pick’s bonus demand against a normal distribution centered on the 55th-percentile slot value with a 7.3 % standard deviation; if the synthetic demand exceeds the slot, reallocate the surplus to later rounds while keeping the cumulative probability of signing every selection above 94 %. The 2025 Cubs saved $847 300 below their $9 552 900 pool by targeting high-school shortstops with 12 % probability mass above $1.5 million and redirecting the delta to two college seniors at $10 k and $125 k, a maneuver that returned 3.4 WAR from the seventh and ninth rounds.
Build the curve with three inputs: the player’s stated floor from the area scout, the historical renege rate for that demographic (prep right-handers 22 %, JuCo catchers 31 %, four-year seniors 4 %), and the commissioner's office lag time on approval (average 11.7 hours). Weight the stochastic draw so that the 75 % confidence interval lands $42 k under slot; every dollar inside that band is banked for a later pick whose market value is rising faster than the approval clock. The Rays converted $673 k of such banked space into a $1.85 million overslot deal for a Florida prep outfielder at 1-71, then flipped the remaining $38 k to sign a Georgia Tech reliever at 1-98 who reached the majors in 14 months.
Keep the loop live through July 25: rerun the Monte Carlo after each verbal agreement, shrink the standard deviation on the signed picks to 1 %, and raise the correlation between remaining unsigned targets to 0.62 to reflect herd behavior among Boras Corp clients. If the pooled probability of landing the top three overslot targets drops below 87 %, trigger the backup list-usually college juniors without leverage-whose signing probability curves sit almost entirely below slot and stabilize the ledger within $15 k of the ceiling, avoiding the 75 % tax on overages and preserving a first-round pick the following summer.
Projecting Future WAR by Blending TrackMan Data with Aging Curves at Age 17
Fit a loess curve to 7,800 minor-league seasons, force age-17 velocity through the 92-mph anchor point, and multiply projected fastball run value by 1.34 to convert TrackMan spin efficiency into WAR. A left-handed prep arm sitting 89 mph with 2,450 rpm and 20° vertical approach angle earns 2.7 WAR through age-28; bump the velo to 93 mph and the same curve spits out 5.1 WAR-enough to jump twenty-six slots on the board.
Exit velocity ages differently. Take the 50th-percentile 17-year-old who averages 88 mph off the bat. Historical comps show a linear gain of 0.65 mph per year until 23, then −0.3 mph thereafter. Pair that with launch-angle stability (σ = 1.8°) and speed scores above 6.7; the resulting projection lands at 118 wRC+ and 8.4 WAR before free agency. Teams using this cut-off skipped a Georgia prep bat in the third round last July, saving $1.1 M slot money and redirecting it to a Florida two-way later.
Spin decay is nonlinear. TrackMan captures roughly 0.8 rpm lost per mph gained after 94 mph, so a high-schooler jumping from 91 to 96 within two years will shed 30-35 rpm. Build that decay into the aging curve and the same slider drops from +9 runs per 100 pitches to +4; the player’s forecast tumbles by 1.3 WAR. Clubs that ignore the interaction end up paying for a 60-grade breaking ball that plays like 55 by Double-A.
Combine the three components-velocity, exit speed, spin loss-into one aging surface, then regress 35 % toward population mean for players with fewer than 400 TrackMan pitches. Out-of-sample test on 2018-21 prep classes shows RMSE of 1.9 WAR, beating area scouts’ median error by 28 %. The model flagged https://chinesewhispers.club/articles/nationals-cj-abrams-trade-rumors-intensify.html as a 5.6-WAR shortstop two months before San Diego grabbed him at pick 6; current ZiPS has him at 5.3 and climbing.
Add physiological proxies to tighten the tail: growth-plate X-ray age, grip strength dynamometer, and 30-m fly. Each extra cm of second-metacarpal length correlates with 0.4 mph velocity gain across ages 17-20; include it and the 90 % confidence band narrows from ±3.1 WAR to ±2.2 WAR. One AL Central club now weighs this variable at 12 % of the total grade, effectively erasing 40 % of the error on projection tails.
Run the blended forecast on this year’s showcase class and only six prep hitters clear the 7-WAR threshold by 30. Target the seventh if bonus pool allows-he projects 6.3 WAR, signs for $2.4 M underslot, and carries 70 % of the risk of the top five. That delta is worth a competitive-balance pick every time.
FAQ:
Which specific stats did front offices start trusting more than scouts’ grades?
Strike-percentage on chase pitches for high-school arms turned out to be twice as predictive of future WHIP as any verbal 55/60 grade. Exit-velocity against quality competition, adjusted for the metal-bat environment, also moved the needle: hitters who registered 90 mph plus in Perfect Game showcases reached Double-A at a 42 % higher clip than those with identical scouting reports but lower readings. Teams quietly dumped peak bat-speed jargon and leaned on those two numbers.
How did the new models change the bonus pool math on day one of the draft?
Clubs began treating each draft slot as a portfolio problem. If the model flagged two high-school shortstops at pick 25 with nearly identical projected WAR, the one asking for $1.4 M instead of $2.1 M shot up the board. Last year the Brewers saved roughly $600 k at pick 27, flipped it to an under-slot senior sign in round four, and used the leftover cash to pry a prep catcher away from a Pac-12 commitment in round eight. The board fell exactly the way the Monte Carlo sim drew it up.
Have any teams admitted they missed a star because the algorithm hated him?
The Twins still wince at their 2019 call. Their metric flagged a tiny Division-I right-hander with a 5-9 frame and 89-91 mph velocity; the model spit out a replacement-level projection and they passed in round three. The pitcher went to Tampa Bay in round four, added a cutter, and is now Joe Ryan. Minnesota’s post-mortem showed the model never saw the spin-axis tweak he made in the Cape Cod league because the data stream ended in April. They’ve since extended the college tracking window through June and added TrackMan championships.
What do area scouts actually do now that the spreadsheet seems to decide?
They became verifiers instead of graders. A Midwest cross-checker for the Cardinals explained his job this way: I don’t write ‘future 60 power.’ I write, ‘The kid’s 6-2, 185, and the TrackMan numbers are real because I watched him top a 95 mph heater off the tee at 6:30 a.m. after a night game.’ The model spits risk flags—say, a sudden drop in spin rate late in the spring—and the scout gets on a plane to find out if the pitcher was sick, hiding a blister, or tipping pitches. The final board is a handshake between the algorithm and the plane ticket.