A predictive layer built on top of Filth+. This document describes how it is constructed, how it was validated, and — just as importantly — what it does not claim. The model exists to forecast how a pitcher's bat-missing stuff ages, with an honest confidence interval around that forecast.
Conventional minor-league aging analysis is built on outcomes — ERA, strikeout rate, swinging-strike rate. In the minors those signals are contaminated, because a 19-year-old in Low-A and a 22-year-old in Triple-A face entirely different competition. A number that moves season to season may reflect the pitcher changing, or merely the league around him changing, and outcome data cannot separate the two.
Filth+ avoids that confound. A pitch's movement profile — velocity, induced vertical break, horizontal break, approach angle, spin, release geometry, and its shape relative to the pitcher's own fastball — is a physical property of the pitch. It is what it is regardless of which level's hitters are standing in the box. Bat-missing ability derives from shape, and shape is level-independent. Aging the grade therefore strips out the single worst confounder in minor-league aging work. We age the pitch, not the box score.
The aging path is learned from the only multi-season pitch-tracking panel that exists at scale: MLB Statcast (Hawk-Eye), 2021–2023 — three full seasons, 1,123 distinct pitchers, collapsed to 7,099 pitcher-season-pitch-type rows after a minimum of 50 tracked pitches of a given type in a season (a stability floor for curve-fitting, distinct from the platform's display floor). This is the same data family Filth+ itself is trained on.
Every pitch in the panel was scored with the production Filth+ model — the identical pickled model, feature pipeline, fastball-differential construction, and global scaling used live on the platform. The aged metric is therefore byte-for-byte the metric shown on player cards, not a re-implementation. As a faithfulness check, the model applied across the full MLB pitch population returns a mean of ~102 and a standard deviation of ~12.8, the expected signature of the production metric applied to all pitches (the scale is anchored to swung-at pitches; applying it to every pitch shifts the mean up a touch by construction — the same modest offset the live grades carry).
Aging is estimated with the delta method (the standard approach in the aging-curve literature), which is specifically chosen to suppress survivorship bias:
This yields 3,224 year-over-year delta pairs — the raw material of the curve.
We model two distinct quantities, because they answer different questions:
Component curves (the narrative layer). Each physical input — velocity, IVB, HB, VAA, spin, extension, release height and side — is aged independently and reported in its native units. This produces the scouting story: how a given pitcher's velocity, ride, and movement are each expected to drift. Components are not recombined into a projected Filth+, because they are physically coupled (IVB declines partly because velocity and spin decline); recomposing from independently-aged inputs would double-count the same underlying change.
The Filth+ curve (the headline). To produce a single defensible projection, the finished Filth+ grade is aged directly with the same delta method — no recomposition, no double-counting. The component curves explain why; the directly-aged Filth+ curve gives the what.
A model that has never been checked against ground truth is an opinion. Before trusting the curve on metrics no one has charted, we validated it against the one metric the public literature has measured thoroughly: fastball velocity aging.
The delta pipeline, run on fastball velocity with no tuning, reproduces the established shape:
The pipeline hit the most-studied figure in the field, out of sample, having never seen that target. That is the validation gate. Having passed it, the same machinery is applied to the components and to Filth+.
The component curves were then checked against physical expectation as a second gate: release geometry, approach angle, and horizontal break are mechanical signatures and should barely move with age — they don't (each drifts under ~0.6 units across a career). Velocity and IVB are the genuine decliners. Spin declines mildly. Nothing moved that physics says shouldn't, and nothing stayed flat that should change. The method does not invent motion where there is none.
The directly-aged Filth+ curve is shallow. Across an entire career it spans only about 2–3 Filth+ points — on a scale where 10 points is a full standard deviation. Aging moves a pitcher's bat-missing grade by at most roughly a quarter of a standard deviation. The curve peaks around age 28 (really a broad plateau from the mid-20s into the early 30s) and declines gradually thereafter.
This is a substantive result, not a null one. Velocity declines meaningfully with age; Filth+ does not, because Filth+ blends velocity with movement, the fastball differential, and release shape — and pitchers adapt, leaning on movement and pitch mix as velocity erodes. Stuff quality is more age-stable than velocity. The honest implication for projection: the aging adjustment is a refinement, not the main event. A pitcher's projected grade is dominated by his current grade; age applies a modest correction. A model claiming age dramatically reshapes stuff would be overstating its hand.
For a pitcher at current age A with current Filth+ F, the projection to a future age is:
projected Filth+ = current Filth+ + (aging curve at target age − aging curve at current age)
The forecast is reported not as a line but as a confidence cone, because a single confident number would be dishonest. Individual pitchers scatter around the population path; we measured that scatter directly. The residual standard deviation of year-over-year Filth+ change, around the population path, is 2.04 points per year. The cone widens with the square root of the projection horizon (random-walk accumulation): roughly ±2.0 points one year out, ±3.5 at three years, ±4.6 at five.
The cone also responds to sample reliability, which falls out of the data: pitchers with deep tracking samples scatter less around the path (residual ~1.85 points) than thin-sample pitchers (~2.69). Established arms get a tighter cone; small samples get a wider one.
A natural challenge: how does this handle a pitcher who gains several ticks in his late 20s and reinvents himself? The honest answer is that no shape-and-age model can predict that spike, because its cause — a mechanical change, a strength or health change, a change in intent — lives entirely outside the data. Velocity and shape do not contain the offseason throwing program that produced the jump.
What the model does instead is more useful than a false prediction: it places that outcome in the upper tail of the cone and quantifies how rare it is. Our own numbers make the point precisely. A one-standard-deviation stuff breakout (10 Filth+ points) over three years sits at roughly 2.8 cone-standard-deviations out — about a 0.3% event. The aging signal across an entire career is ~0.25 SD; a breakout of that magnitude is multiples of the entire aging range. The data itself confirms that such jumps are driven by non-age factors and are not forecastable from shape and age. The model's job is to draw the envelope and mark where the tail lives — never to pretend the tail is the trend.
The decisive next step is a public prediction-validation record: timestamping projections now and grading them as the season unfolds. The aging curve will also become natively empirical as the platform's own multi-season tracked data accumulates — a minor-league Filth+ aging curve built on observed development rather than transferred from the majors is a two-to-three-season horizon, and the collection begins now.
Methodology references the Stuff+ family of shape-based pitch grading as the originating public approach. The aging method follows the standard delta-and-chaining framework for aging curves. All figures above are from the model as described and are reproducible from the panel.