The DMX Methodology Whitepaper

DynastyFootballFan · June 2026 · v1.0

Abstract

This paper documents the full methodology behind DMX (Dynasty MetriX), a pre-draft prospect score for dynasty fantasy football, and its in-career counterpart DPX (Dynasty Performance Index). It covers the 25-year dataset the models are trained on, the component architecture, how career outcomes are defined and labeled, the complete published validation results (12 regressions across 4 positions and 3 target metrics, plus decile hit rates on 1,853 completed careers), and an honest account of known limitations. Three categories of model internals are deliberately withheld as trade secrets; everything else — including every number needed to independently verify the model's predictive claims — is published here and on the live platform.

1. The problem: pricing rookies before the league does

Dynasty leagues are decided at the moments of maximum uncertainty — rookie drafts and the trades around them — where the market runs on draft-media narrative, recency, and film takes of wildly varying quality. The question DMX answers is narrower and more disciplined than “who is good?”: given only the information that exists on draft night, what did historically comparable prospects actually do in the NFL?

That framing has two consequences. First, the model must be built the same way for every prospect in every class, or its historical hit rates mean nothing. Second, it must be evaluated against real career outcomes over a window long enough to label them — which means a model without many years of history cannot have a track record, no matter how clever its features. The dataset, not the formula, is the foundation.

2. The dataset

The training and validation corpus covers 3,187 offensive prospects across 27 NFL draft classes (2000–2026) and 2,987 defensive (IDP) prospects since 2004. Each prospect is identity-matched across four domains that public football data does not join cleanly on its own:

Most of the engineering effort lives in the joins. Public football data is riddled with duplicate identities, name collisions, transfers, and inconsistent IDs across providers; the platform maintains a canonical player-identity spine with alias resolution so that the same human is the same row in every table. A scoring model built on dirty joins doesn't fail loudly — it quietly scores the wrong careers, which is worse.

Gaps are kept visible rather than imputed away. Comprehensive defensive college statistics do not exist in any clean public source before 2004, so early IDP classes carry reduced-confidence scores with the production component explicitly null rather than silently zeroed (see §7).

3. DMX architecture

DMX is a composite z-score built from three orthogonal dimensions of pre-draft information, each expressed within the prospect's position cohort:

3.1 Athleticism (ATH)

Combine and pro-day testing: 40-yard dash, vertical jump, broad jump, 3-cone, shuttle, bench press, and height-adjusted weight. Each metric is z-scored within position, then rolled into four published sub-components — Speed (ATH-SPD), Agility (ATH-AGIL), Lower-body Power (ATH-LP), and Strength (ATH-STR) — and finally into a single ATH composite.

3.2 Draft Position (DPOS)

Pick number transformed to a z-score within the position cohort, with undrafted players assigned pick 300. DPOS is the NFL's aggregate scouting consensus expressed as one number, and it is the strongest single-variable predictor at every position. Any prospect model that ignores draft capital is throwing away its best feature; any model that uses only draft capital is leaving measurable accuracy behind (see §5).

3.3 Age-Weighted Production (AWP)

Final college season per-game production, normalized by age and competition tier (Power 5 / Group of 5 / FCS). Sub-signals include market-share yardage, age-relative-to-class, an early-declare “phenom” signal, and dominator-style shares of team output. For IDP prospects, AWP is computed as havoc share — the player's share of team defensive disruption (sacks + TFL + forced fumbles + INT + passes defensed) — which travels across schemes better than raw counting stats.

3.4 Composite assembly and deciles

Per-position weights blend the three components into the DMX composite. The weights are calibrated by maximum-likelihood fit against career outcomes on the 25-year training set and refit roughly every three years; the specific values are not published (§8). The familiar D1–D10 labels are then assigned per (position, draft-year) cohort — NTILE(10) OVER (PARTITION BY position, draft_year ORDER BY dmx DESC) — so a decile is always relative to that year's class. The raw score is the better tool for cross-class comparison.

Deliberately not in DMX: landing spot (the score is computed at the moment of the draft), injury history as a separate signal (it is visible to NFL scouts and therefore already priced into DPOS), and character/interview information (signal-poor, noise-rich).

4. Defining career outcomes

A prospect model is only testable if “success” is defined before looking at the predictions. Three target metrics are used:

Labeling requires waiting: a 2021 draftee's 5-year window only closed after the 2025 season. This is why the validation cohorts below end at 2020–2023 even though scoring runs through the 2026 class.

5. Validation

5.1 The complete regression table

The table below is the full output of all 12 active univariate regressions of DMX against career outcomes — 4 positions × 3 targets, fit on every prospect with a complete career window. Nothing is cherry-picked: every active regression in the database is shown, and the same values are queryable live on the Model Transparency page.

PositionTargetInterceptSlopeNCohort
QB5-yr VBD16.536.00.1473162000–2020
RB5-yr VBD26.052.10.1647011998–2020
TE5-yr VBD11.927.30.0845101997–2020
WR5-yr VBD20.942.30.10510881996–2020
QBPeak 3-yr VBD avg8.614.50.0781462000–2020
RBPeak 3-yr VBD avg9.218.80.1513511999–2020
TEPeak 3-yr VBD avg4.512.10.0982362001–2020
WRPeak 3-yr VBD avg10.915.50.0724592001–2020
QBCareer tier (ordinal)1.50.80.2462452000–2023
RBCareer tier (ordinal)1.80.60.2225771998–2023
TECareer tier (ordinal)1.70.80.3264132001–2023
WRCareer tier (ordinal)1.70.60.1928032001–2023

predicted_outcome = intercept + slope × DMX. Active regressions from dmx_model_coefficients, June 2026.

Two honest readings. First, R² in the 7–33% range is what real pre-draft prediction looks like: coaching changes, scheme fit, injuries, and depth-chart competition are invisible on draft night, and they cap what any model can know. Treat the remainder not as noise to dismiss but as a quantified ceiling on draft-day knowledge. Second, the tier regressions run roughly 2× the R² of the continuous VBD regressions — career outcomes cluster into archetypes far more cleanly than they distribute along exact point totals, and “which archetype” is the more learnable (and for roster decisions, more useful) question.

5.2 Decile hit rates

Collapsing to the score's most common consumption format — deciles — here is what every offensive prospect from the 2001–2020 classes with a completed career window actually became:

DecileProspectsStar %Star or Starter %Bust %
D123737.6%54.9%14.3%
D222925.8%47.6%21.4%
D321020.0%38.1%35.2%
D422510.2%22.2%49.8%
D51656.1%18.8%57.0%
D61444.9%16.7%71.5%
D71727.0%12.8%69.2%
D81791.7%10.6%73.7%
D91482.7%10.1%75.0%
D101441.4%5.6%79.9%

n = 1,853 offensive prospects, draft classes 2001–2020, career tier labeled.

The gradient is monotonic in the direction that matters — the top decile produces a Star or long-term Starter at roughly 10× the rate of the bottom three. The same cut by position, comparing the top two deciles against the bottom two:

PositionD1–D2 hit ratenD9–D10 hit raten
QB57.9%577.7%26
RB53.3%1358.6%93
TE55.4%923.2%62
WR45.6%1829.9%111

Hit = career tier of Star or Starter. Same 2001–2020 labeled cohort.

5.3 Calibration and residuals

Beyond rank ordering, the model is checked for calibration — whether predicted outcomes match actual outcomes on average within each prediction bucket. Per-position calibration plots (every completed-career prospect, prediction vs. actual) are published interactively on the Model Transparency page, and ranked over/under-performer residual leaderboards are on the Analytics page. The residuals are a feature, not an embarrassment: the players the model missed, and the directions it missed in, are exactly where the next refit learns.

6. DPX: the in-career model

DMX freezes at the draft by design. From a player's first NFL season onward, dynasty value is tracked by DPX, a separate model built on three observable career components, each a z-score updated from real game data: Volume (touches, targets, snap share), Scoring (fantasy production per game, half-PPR and PPR pathways — the heaviest-weighted layer, since dynasty value ultimately compounds from scoring), and Efficiency (production per opportunity). A dynasty modifier reflecting contract status and draft-capital decay sits on top, separating dynasty value from redraft value.

DPX publishes two deciles per player-season: DPX-Next (next-season projection) and DPX-Long (three-year forward projection). Thirteen NFL seasons of player-scores are maintained, with roughly 350 active players scored per season. The full consumer-level explainer, including a worked example and reading patterns, is at What is DPX?.

7. Known limitations

8. What we don't publish, and why

Three categories of model internals are withheld: (1) the DMX position-component weights, (2) the IDP DMX position-component weights, and (3) the DPX layer weights and dynasty-modifier coefficients. They are the product of 25 years of iterative refinement on a curated proprietary dataset and constitute the platform's commercial differentiation. Anyone can average three z-scores; which blend best predicts NFL outcomes is the hard-won part.

The disclosure principle: outputs are public, internals are not. Every score, decile, sub-component, regression coefficient, and hit rate in this paper is publicly readable from the live database; every weight that produces them is locked behind row-level security with anonymous access revoked.

9. Reproducibility

The validation claims in §5 do not require trusting us. DMX scores for completed-career cohorts are publicly readable; career VBD and tiers can be computed independently from public play-by-play data (e.g. nflverse); OLS on those pairs should land within rounding of the coefficients published above. If it doesn't, that's a model bug we want to know about — the contact path is the newsletter reply address.

For the consumer-level introductions to both scores, see What is DMX? and What is DPX?. For the live interactive version of every table in this paper, see Model Transparency.