MLB Pitcher Velocity Trend Model: A Smarter Way to Track Fastball Trends and Spot Risk Early

Posted Feb. 25, 2026, 9:08 a.m. by Lesly Shone 1 min read

Velocity reveals the truth about a pitcher’s body, workload, and command more reliably than any surface stat. The MLB pitcher velocity trend model is designed to track fastball speed over time, predict next-start velocity, identify risky dips, and take into account context like rest, weather, and ballpark effects. It provides actionable insights for bettors, analysts, and fantasy managers. This model is practical, data-driven, and structured to help users make informed decisions across games without relying on hype or guesswork.

Table of Contents

Purpose and Outcome
Data and Sources
Feature Engineering
Modeling Approach
Validation and Operations
How Bettors Can Use It on ATSWins
Step-by-Step Build Instructions
Practical Modeling Tips
Example Alert Rulebook for Operators
Dashboard Elements That Help Decision-Making
Conclusion
Frequently Asked Questions (FAQs)

Key Takeaways

Velocity trends are an early signal for performance and risk. By predicting next-start fastball speed, tracking short-term slopes, and estimating the chance of a meaningful drop, users can adjust K props, outs, or runs. Clean Statcast data, paired with rest days, pitch counts, weather, ballpark, and opponent quality, forms the foundation. Building steady features like rolling 3-, 5-, and 10-start windows, using EWMA smoothing, and splitting first-inning versus later-inning velocity improves signal quality. Smoothed predictions through a Kalman filter, combined with a tree booster or GAM, allow accurate next-start estimates. ATSWins integrates these signals to offer data-driven picks, player props, betting splits, and profit tracking across MLB, NFL, NBA, NHL, and NCAA, helping bettors act on edges that matter.

Purpose and Outcome

The MLB pitcher velocity trend model estimates a pitcher’s true fastball velocity in near real time. It produces three key outputs before each scheduled start. First, it forecasts next-start velocity for the primary fastball. Second, it calculates the short-term slope to show if the pitcher’s velocity is climbing, holding steady, or dropping. Third, it estimates the probability of a meaningful drop, typically greater than 1.5 miles per hour, which could indicate strain, fatigue, or potential injury.

Pitcher velocity is one of the most actionable signals for predicting command, swing-and-miss likelihood, and overall performance stability. Small shifts in velocity often correlate with whiff rate, strikeout potential, and weak contact. Even a one-mile-per-hour drop can affect expected strikeout percentage, walk rate, and home run risk. Short rest periods, high pitch counts, and recent high-leverage innings often precede minor dips that are significant for betting or fantasy analysis. Sudden velocity and spin losses frequently precede injury stints, making early identification critical.

This model relies on consistent Statcast baselines from Baseball Savant and incorporates practical insights from pitching development practices. By focusing on actual pitcher behavior instead of one-off narratives, it avoids overfitting to hype or anecdotal reports. The goal is to detect meaningful trends before the broader market reacts.

The target outputs include next-start velocity as a point forecast with prediction intervals, the slope of the short-term trend using a rolling three-to-five game change plus EWMA smoothing, and the risk of a velocity drop relative to the pitcher’s baseline over the past ten appearances. Alerts are categorized into amber and red. Amber signals occur when projected drops reach one mile per hour or when a one-and-a-half-mile-per-hour drop happens in the last start without a spin decrease. Red alerts trigger when projected drops reach 1.5 miles per hour with a concurrent spin drop or when a two-mile-per-hour drop occurs alone.

For bettors on ATSWins, these signals have multiple applications. Player props, including strikeouts, outs recorded, and earned runs, can be adjusted based on predicted fastball quality and trend slopes. A one-to-one-and-a-half-mile-per-hour drop can translate into an expected difference of 0.3 to 0.7 strikeouts per start, depending on the pitcher archetype. Sides and totals are influenced as velocity and spin stability tighten expected outcome distributions, while elevated drop risk can expand tails, nudging totals. In live betting, first-inning velocity checks allow adjustments to moneylines and totals. Season-long angles, such as aging curves and cumulative fatigue, help identify fade spots and improve hierarchical pooling for rookies or post-injury returns.

Data and Sources

The core datasets include pitch-level Statcast data collected via Baseball Savant or programmatically through pybaseball. Important fields include release speed, spin rate, pitch type, zone, stance, game date, inning, pitcher, batter, and events. Game logs provide context such as rest days, pitch count, batters faced, and high-leverage innings. Park and weather variables include park ID, park factor, temperature, humidity, wind direction and speed, and roof state. Opponent quality incorporates rolling metrics like wRC+, strikeout percentage, swing tendencies, and handedness splits from FanGraphs.

Data is pulled from the last three to five seasons. Each pitcher-date is aggregated to capture maximum, median, and 95th percentile velocities for each fastball type. Game conditions and opponent quality are tagged. Pitcher roles, such as starter, bulk reliever behind an opener, or pure reliever, are recorded. Partial outings are flagged but included for slope estimation with appropriate uncertainty.

Cleaning involves removing outliers beyond four standard deviations from the median within pitch type, winsorizing spin-rate extremes, and unifying pitch type taxonomy. Ambiguous fastballs, like four-seam versus cutter, are consolidated based on seam shift and seasonal frequency. Openers and partial outings are included with down-weighting for baselining. Weather data is normalized and merged with park factors, using medians to backfill missing values.

Minimal schema columns include pitcher ID, hand, age, game date, team, opponent, park, role, pitches thrown, batters faced, first-inning flag, pitch type, median and max release speeds, median spin, release position variance, rest days, travel flags, temperature, humidity, wind speed, roof state, opponent quality metrics, and injury flags.

Feature Engineering

Feature engineering transforms raw Statcast and context data into meaningful signals for predicting velocity trends. The foundation begins with rolling and smoothing metrics for each pitcher and fastball type. Rolling medians are calculated across three-, five-, and ten-game windows. These capture short- and medium-term velocity patterns, reducing the impact of one-off fluctuations. Alongside medians, the maximum velocities are tracked. Maximum speed reflects top-end effort or intent, while the median captures cruising velocity. The spread between the two can indicate a pitcher’s ability to reach back in crucial counts or hint at fatigue when top-end velocity drops disproportionately.

To reduce noise and highlight recent trends, exponentially weighted moving averages are applied to median velocities, typically with an alpha between 0.3 and 0.5. This approach allows the model to respond to meaningful changes without overreacting to a single bad outing. All rolling and smoothed features are computed by sorting appearances chronologically for each pitcher and fastball type, producing median and maximum values across three-, five-, and ten-game windows and EWMA versions.

Game-state splits further enhance the model by separating first-inning velocity from later innings. First-inning medians capture pitchers’ initial effort, often influenced by warm-up or adrenaline, while later-inning medians reflect pacing, endurance, and mechanical consistency. The difference between these two, called the delta-first-later, can signal early signs of conditioning issues or fatigue if the first-inning velocity is notably higher than subsequent innings.

Comparisons against a pitcher’s own career or season baseline provide additional context. For each pitcher, a career baseline is established after a minimum of 200 fastballs thrown. Z-scores are then computed for rolling ten-game median velocities relative to this baseline. A z-score between -1.0 and -1.5 signals a yellow alert, highlighting a significant drop relative to that pitcher’s typical performance.

Slope and change-point features capture abrupt shifts and trends. A simple linear regression is fit over the last five appearances for the EWMA median fastball velocity, yielding a slope in miles per hour per start. Change-point flags identify meaningful deviations by comparing the most recent two appearances to the prior five using a pooled t-test or Bayesian change detection. Flags are triggered when the level shift exceeds a defined threshold, such as 0.8 miles per hour with a high probability. These signals help anticipate sudden drops in performance before they fully manifest in traditional stats.

Spin-rate metrics and release-point stability complement velocity tracking. Median spin over the last three games and deltas between three- and five-game windows are calculated. Rolling standard deviations for release positions in both horizontal and vertical axes provide a measure of mechanical consistency. Rising variance often corresponds with fatigue or mechanical adjustments. A combined signal of spin drop and rising release variance is a stronger indicator of risk than either factor alone.

Pitch-mix shifts are also relevant. The share of fastballs thrown over the last three games is compared to season-to-date usage. A spike in offspeed pitches alongside fastball velocity drops can either reflect strategic sequencing or indicate a red flag. Including both the mix shift and its interaction with velocity slope allows the model to capture context-dependent effects.

Environmental adjustments account for temperature and park factors. Historical velocities are regressed on temperature, often using spline regression to handle non-linear effects. Observed velocities are then residualized by subtracting the expected temperature effect, producing features for raw and adjusted velocities. Park factors are included to capture idiosyncratic measurement effects.

Rest and workload indicators further refine predictions. Rest-day features include categorical bins, such as three or fewer, four, five, or six or more days, with non-linear splines for exact day counts. Short rest markers highlight back-to-back appearances with fewer than four rest days, high pitch counts exceeding 100 in the previous start, or consecutive high-leverage appearances from bullpen to starter. Cumulative stress measures like total pitches thrown over the last 30 days and maximum single-inning counts inform fatigue-sensitive adjustments.

Alerts are generated by combining these signals. A red alert is triggered when the last two starts show a velocity drop of at least 1.5 miles per hour versus the ten-game EWMA baseline, coupled with a spin drop of 80 revolutions per minute or more, or when a projected drop exceeds two miles per hour. Amber alerts flag projected drops above one mile per hour or when the change-point probability exceeds 0.7 with a negative slope, even if spin is stable.

Adjustments for handedness, age, and opponent characteristics are also incorporated. Opponent platoon splits are applied to reflect the projected right- and left-handed batter mix. Pitchers in their mid-thirties often experience sharper declines after high-stress outings, so age-related non-linearities are embedded in the model through splines or hierarchical priors.

These features collectively form a rich representation of each pitcher’s short-term condition, mechanical stability, and environmental influences. By carefully combining rolling medians, EWMA smoothing, game-state splits, z-scores, slope and change-point flags, spin and release variance, pitch-mix shifts, environmental corrections, rest markers, and contextual adjustments, the model can robustly predict next-start velocity and identify high-risk dips. This structured approach transforms noisy, complex data into actionable signals for ATSWins users seeking an edge in betting, fantasy, and analytical evaluation.

Modeling Approach

Once the feature layer is stable, the mlb pitcher velocity trend model moves into estimation. The central idea is simple. Observed game level velocity is noisy. True underlying arm strength and mechanical efficiency evolve more smoothly. The modeling stack separates those two layers, then reconnects them to forecast the next start.

The first component is latent velocity smoothing. Observed appearance medians fluctuate because of weather, opponent approach, pitch mix, and even random measurement error. Treating each start as the truth creates overreaction. Instead, a state space structure estimates an unobserved true velocity that evolves over time. The state equation allows true velocity to drift gradually from one appearance to the next. The observation equation links the hidden state to observed medians with measurement noise. A Kalman filter recovers a smoothed trajectory that reflects the underlying direction rather than random bumps.

This structure handles partial outings naturally. A twenty-five pitch opener appearance produces higher uncertainty than a ninety pitch start. The filter accounts for that through observation variance. Context features such as temperature and rest days can enter as exogenous drivers of either state drift or observation adjustment. That flexibility matters early in the season when cold conditions artificially depress readings.

The second layer introduces hierarchical pooling. Not every pitcher has five seasons of stable data. Rookies, call-ups, and post-injury returns need prior information. A hierarchical Bayesian framework models baseline velocity as partially shared across groups. Age bands, handedness, and pitch family archetypes define grouping structure. A thirty-four-year-old right-handed four-seam heavy pitcher shares more prior information with similar profiles than with a twenty-three-year-old sinker specialist. This partial pooling stabilizes forecasts when sample sizes are thin.

The practical implementation can be built with tools like PyMC for Bayesian estimation or approximated with empirical Bayes shrinkage inside a frequentist workflow. Full Bayesian estimation is slower but yields direct uncertainty intervals. For operational pipelines, weekly re-estimation balances computational cost and stability.

After latent velocity is estimated, a supervised mapping predicts the next start velocity. This regression model takes smoothed current velocity, short-term slope, spin deltas, release variance, rest bins, pitch count stress, temperature forecast, park context, and opponent metrics as inputs. Two common choices work well. A generalized additive model captures nonlinear relationships transparently. Gradient boosting models like XGBoost or LightGBM capture interactions and complex patterns efficiently. In practice, ensembling both often improves stability.

The regression output produces a point forecast and interval band for next start fastball velocity. That feeds directly into betting adjustments. Alongside regression, a classification head estimates the drop risk probability. The target variable equals one when next start velocity lands at least 1.5 mph below the exponentially weighted baseline. Because true large drops are relatively rare, class imbalance techniques such as weighted loss or focal loss help maintain sensitivity without flooding alerts.

Calibration is mandatory. Raw classifier probabilities often drift. Isotonic regression or Platt scaling aligns predicted probabilities with empirical frequencies. A drop probability of 0.60 should mean roughly sixty percent of similar cases historically experienced that magnitude of drop.

Leakage avoidance remains critical. All features must be frozen as of the last completed appearance. Opponent metrics use only data available before the upcoming start. Weather inputs use forecasted values when simulating pre-game decisions. In backtests, actual weather can evaluate forecast quality, but never train with information that would not have existed at prediction time.

Interactions capture nuance. Pitcher handedness interacting with opponent projected right-handed share influences pacing. Temperature interacting with the roof state captures indoor stadium moderation. Velocity slope interacting with spin drop amplifies red flag signals. Tree based models learn many of these automatically, but sanity checking partial dependence plots prevents unrealistic relationships.

Validation and ops

A model that cannot survive walk-forward validation does not belong in production. Time-based splits simulate real-world deployment. Training uses data up to a cutoff date. Predictions are generated for subsequent starts. The cutoff moves forward incrementally. This approach mimics how the model would have behaved historically rather than benefiting from hindsight.

For velocity regression, mean absolute error and root mean squared error measure accuracy. Coverage of fifty percent and eighty percent prediction intervals checks uncertainty calibration. For drop risk classification, receiver operating characteristic area under the curve and precision recall area under the curve quantify discrimination. Brier score measures probability accuracy. Reliability diagrams visualize calibration.

Monitoring extends beyond global metrics. Monthly breakdowns by temperature regime reveal seasonal drift. Early April cold games often produce slight systematic underpredictions if temperature normalization undercorrects. Mid summer heat can produce the opposite effect. Drift detection thresholds trigger re-tuning when the error increases materially.

Threshold tuning for alert tiers balances sensitivity and noise. False negatives risk missing valuable betting edges and early injury signals. False positives clutter dashboards and reduce trust. Cost-aware optimization considers expected betting value per alert. Amber tier might activate when drop probability exceeds 0.35 or the projected drop exceeds 1.0 mph. Red tier may require drop probability above 0.55, combined with projected drop above 1.5 mph or confirmed spin decline.

Deployment follows a nightly batch structure. Data refresh pulls the latest Statcast entries from Baseball Savant through pybaseball. Feature pipelines recompute rolling windows and smoothing inputs. Lightweight regression layers refit with warm starts. Weekly, the hierarchical prior and state noise parameters are refreshed. Predictions are published to the ATSwins dashboard, attaching projected velocity, slope, drop probability, and alert color to each probable starter.

Monitoring dashboards track model error, alert frequency, and realized betting performance. If model error increases by twenty percent month over month, investigation begins immediately. Data quality checks verify pitch type tagging stability and weather joins. Operational discipline protects the edge.

Human oversight remains part of the system. Beat reports of mechanical adjustments or velocity building programs influence uncertainty widening for two starts. Role transitions from bullpen to rotation increase process noise temporarily. Portfolio controls limit exposure when alerts rely solely on small sample slopes without corroborating spin or release signals.

How Bettors Can Use It On ATSwins

ATSwins integrates the mlb pitcher velocity trend model directly into its MLB workflow. On the MLB games page, available at ATSwins, probable starters display projected velocity, short-term slope, and drop risk indicator. The workflow begins with scanning for amber and red tier signals. A steep negative slope of minus 0.6 mph over the last five appearances draws attention even before drop probability crosses a threshold.

Strikeout props respond quickly to velocity changes. A projected drop of 1.2 mph against a contact-heavy opponent justifies a downward adjustment of expected strikeouts. Outs recorded props shift when efficiency decreases due to weaker fastball life. Game totals expand slightly when variance increases. A stable or rising velocity profile tightens expectation and supports unders when other conditions align.

Cross signal confirmation improves confidence. A spin drop of eighty rpm paired with a release variance increase strengthens fade cases. Conversely, a mild projected drop on a hot day in a closed-roof stadium may partially offset physical decline. Temperature forecasts and park context matter.

Live betting introduces another layer. When first inning median velocity lands 1.5 mph below the forecast, immediate reevaluation occurs. If the dip persists into the second inning alongside command inconsistency, live overs or strikeout unders gain value before books fully react.

Season long trends also surface. Aging pitchers accumulating heavy early-season workloads may show gradual slope decline by mid summer. Rookie call ups benefit from hierarchical pooling but remain high uncertainty cases for the first few starts. Tracking realized drops against predicted probabilities builds confidence in system reliability.

All projections, alerts, and outcomes feed into ATSwins profit tracking. Transparency matters. If red alerts historically produce positive return on investment in strikeout unders, unit sizing can scale appropriately. If amber alerts underperform in certain parks, thresholds adjust. Data informs discipline.

Step-by-step Build Instructions

Building the mlb pitcher velocity trend model from scratch requires structured stages. First, extract pitch level Statcast data for at least three to five seasons using pybaseball connected to Baseball Savant. Pull release speed, spin rate, pitch type, release coordinates, game date, and identifiers.

Second, join game logs to compute rest days, pitch counts, and batters faced. Merge park and weather datasets keyed by stadium and date. Integrate opponent rolling metrics from FanGraphs calculated as of the game date to prevent leakage.

Third, clean and standardize pitch types. Map FF, SI, FT, and FC into consolidated hard fastball families where appropriate. Remove velocity outliers beyond four standard deviations per appearance. Winsorize spin extremes.

Fourth, compute rolling medians for three, five, and ten appearance windows. Calculate exponentially weighted moving averages with alpha near 0.4. Derive slope features over last five appearances. Calculate spin and release variance deltas.

Fifth, normalize velocity for temperature effects using regression residuals. Retain raw velocity as well. Construct rest day bins and pitch count stress markers.

Sixth, implement latent smoothing via a Kalman filter or similar state space model. Estimate process and observation noise through backtesting. Output smoothed current velocity and uncertainty.

Seventh, train regression and classification heads using time based splits. Avoid random cross-validation that mixes seasons. Record mean absolute error and probability calibration metrics. Tune hyperparameters carefully, preferring stability over marginal gains.

Eighth, define alert thresholds through expected value optimization rather than arbitrary cutoffs. Backtest betting strategies tied to alerts using historical market lines where available.

Ninth, deploy nightly batch updates and weekly deeper refreshes. Publish predictions to the ATSwins dashboard. Monitor drift and recalibrate as necessary.

Practical Modeling Tips

Fastball identification can drift across seasons. A cutter thrown at four seam velocity may become the primary hard pitch. Establish season dominant pitch type rules rather than game by game classification. Map velocity changes to strikeout effects using within pitcher regressions rather than league wide averages. Power four seam pitchers often experience larger strikeout declines per mph drop than sinker specialists.

Small sample caution cannot be overstated. Two recent starts do not define trend. When only one or two appearances exist for a rookie, widen prediction intervals and shrink the slope toward zero. Require corroborating signals before issuing red alerts in tiny samples.

Nonlinearity in rest days and temperature should be explicitly modeled. Recovery benefits plateau beyond five days. Excessively long rest sometimes reduces sharpness slightly. Temperature effects differ across pitchers. Some lose carry in extreme heat. Partial dependence plots verify realistic relationships.

Weekly sanity checks protect integrity. Plot observed versus predicted velocity for top inning pitchers. Examine residual distributions. Compare predicted drop probabilities against realized frequencies. Investigate systematic bias in specific parks or months.

Example Alert Rulebook For Operators

Amber tier signals moderate caution. Consider fading strikeout props when opponent contact rates are high and park favors hitters. Evaluate leaning toward overs if bullpen depth behind the starter is thin. Red tier signals stronger lean. Strikeout unders and outs recorded unders gain weight. Team total overs against the affected starter become attractive in warm environments. Unit sizing must remain disciplined and grounded in historical return on investment metrics rather than emotional reactions to color coding.

Dashboard Elements That Help Decision-Making

A clear dashboard accelerates workflow. Each pitcher card displays projected next start velocity with fifty and eighty percent bands, short term slope, and drop probability. The last five appearances show median and maximum velocity with spin overlay. Change point flags highlight abrupt shifts. Context panel lists rest days, pitch count, temperature forecast, park, and opponent handedness projection.

Hover interactions reveal top contributing features based on SHAP analysis for tree models. For example, short rest may subtract 0.55 mph while warm temperature adds 0.30 mph. Exportable CSV fields allow integration with personal betting logs.

Conclusion

Velocity trends say a lot about performance and underlying risk. When fastball speed holds steady, it usually supports stable command, normal strikeout rates, and deeper outings. When it slips across multiple starts, that is often an early warning that something is changing, whether it is fatigue, mechanical drift, or role adjustment. The edge comes from identifying those shifts before they fully show up in ERA or surface stats. Clean Statcast data, build rolling windows, smooth the noise, and project next start velocity along with realistic drop probabilities. Layer in rest days, temperature, and spin variance so the projection reflects real context instead of raw averages. The goal is simple: react early instead of chasing what already happened.

ATSWins applies this velocity framework inside a broader AI-driven system that delivers data-backed picks, player props, betting splits, and transparent profit tracking across NFL, NBA, MLB, NHL, and NCAA. The platform blends velocity signals with matchup data and risk filters so bettors can see where stability supports an over or where a downward trend supports an under. With both free and paid options available, users get structured insights without fluff. Track weekly, set alerts for meaningful velocity shifts, and position ahead of the market instead of adjusting after it moves.

Frequently Asked Questions

What is an MLB pitcher velocity trend model?

An MLB pitcher velocity trend model tracks how a pitcher’s fastball speed changes over time and uses that pattern to estimate what the next start might look like. Instead of staring at one box score, it looks at rolling averages, short-term slopes, rest days, pitch counts, and even weather to see whether a pitcher’s arm is stable, building, or quietly dipping. The goal is not to predict injuries. It is to flag meaningful velocity movement before the betting market fully reacts. When velocity trends shift, strikeout upside, command consistency, and outing length can shift with it.

Why does fastball stability matter for betting and the health context?

Fastball stability usually signals mechanical consistency and a fresh arm. When velocity holds steady or trends slightly up, it often supports strikeout props and deeper outings. When it trends down across multiple starts, that can point to fatigue or mechanical drift, from a betting angle, that matters for strikeout props, outs recorded, and even totals if a starter is unlikely to go deep. From a health lens, no diagnosis is ever assumed, but repeated dips paired with shorter outings are yellow flags that should not be ignored.

How can someone build a simple version without overcomplicating it?

Start basic. Track average fastball velocity by game. Create rolling 3, 5, and 10 game averages to see short and medium trends. Add rest days and previous pitch count to account for recovery. Note the temperature, since colder conditions can slightly suppress velocity. Compare each start to the pitcher’s own season baseline instead of league averages. A simple next start estimate can combine baseline velocity with short-term trend and a small rest adjustment. It does not need complex math to be useful.

What inputs matter most, and what can wait?

The most important inputs are recent velocity windows, rest days, pitch count from the last outing, and basic weather context. Spin rate consistency and release point stability can add signal once enough data is available. Early on, avoid overreacting to single-game outliers caused by weather or unusual mound conditions. Also, avoid heavily modeling brand new pitch types until a real sample builds. Keep the structure clean and expand gradually.

How does ATSWins use this model in practice?

ATSWins integrates an MLB pitcher velocity trend model into its broader MLB workflow to assess next start stability and short-term fatigue signals. Those velocity insights are combined with matchup context and risk checks to support picks and player prop evaluations. Results are tracked within the platform so users can see what actually performs over time. The focus stays on clear, data-driven signals that help bettors make sharper decisions without unnecessary noise.

AI Football Betting Tools - How They Make Winning Easier

Bet Like a Pro in 2025 with Sports AI Prediction Tools

Sources

The Game Changer: How AI Is Transforming The World Of Sports Gambling

AI and the Bookie: How Artificial Intelligence is Helping Transform Sports Betting

How to Use AI for Sports Betting