DynastyFootballFan

Model Transparency

How the DMX Scouting Model and DPX Dynasty Model work, what they predict, where they fail, and what we don't publish. Honest accounting of a 25-year dataset rather than marketing claims.

Active Models
Best R²
Training Obs.
27
Draft Classes
1996+
Coverage
What this page is for — If you're going to act on a model, you should understand how it works, what it gets right, and where it breaks. This page documents both. It does not publish the proprietary weights inside DMX, but it does publish every regression coefficient that maps DMX to predicted career outcomes — so you can independently verify the model's predictive claims against your own analysis.
DMX — Dynasty MetriX (Pre-Draft Scouting Model)Composite

DMX is a single composite z-score evaluating NFL Draft prospects at the moment of the draft. It blends three orthogonal dimensions of pre-draft information into one number per player. D1 (top decile) is the elite tier; D10 is the bottom. The composite is computed per position with position-specific weights.

Athleticism
ATH
Combine + pro-day testing: speed, agility, power, strength. Position-normed z-scores blended into a single ATH composite. Sub-components (ATH-SPD, ATH-AGIL, ATH-LP, ATH-STR) are visible on player detail panels.
Draft Position
DPOS
Draft capital z-score within position cohort. Aggregate NFL scouting consensus expressed as a single number. The strongest single-variable predictor at every position (R² 5.6–13.4% standalone).
Age-Weighted Production
AWP
College production normalized by age and competition level. For IDP prospects, this is computed as havoc share — the player's share of team defensive disruption (sacks + TFL + FF + INT + PD).

Position-specific weights are not published — they're the result of 25 years of iterative refinement and are central to what makes DMX distinct from simply summing combine metrics. See the Methodology tab for what this means in practice.

DPX — Dynasty Performance Index (In-Career Performance Model)Composite

DPX tracks dynasty value during an NFL career — orthogonal to DMX, which only predicts pre-draft. Built on three observable career components updated every season:

Volume
35% weight
Touches, targets, snap share. The opportunity baseline — high-volume players sustain value even with average efficiency.
Scoring
40% weight
Actual fantasy production per game. Half-PPR and PPR pathways. Single largest weight because dynasty value ultimately compounds from scoring.
Efficiency
25% weight
Production per opportunity. Catches DPX value players whose volume is volatile but whose per-touch output suggests sustained quality.

DPX is published as two outputs per player-season: DPX-Next (next-season projection decile) and DPX-Long (3-year forward projection decile). Both modified by contract status and draft capital decay.

How predictive is DMX, exactly? The table below is the complete output of 12 univariate regressions of DMX against career outcomes — 4 positions × 3 target metrics — fit on every prospect with a complete career window. Every row is an independent test. Coefficients, R², training size, and cohort years are all published. The active regressions are what drives the predicted_vbd values on the Analytics page residuals tab.
All Active Regressions predicted_outcome = intercept + slope × DMX
Position Target Metric Intercept Slope N Cohort Notes
Loading regressions…
R² by Position & Target% of career outcome variance explained by DMX

Why tier R² runs 2× higher than VBD R²: Career outcomes cluster into discrete archetypes (Star / Starter / Contributor / Bust) much more cleanly than they distribute along continuous fantasy points. Predicting "what archetype" a prospect becomes is a more learnable problem than predicting their exact career point total. Both regressions are honest; the tier framework is more useful for dynasty roster decisions.

Interpretation guidance. R² in the 10–30% range is normal — even strong — for projecting NFL outcomes from college data. NFL outcomes depend heavily on factors that are invisible at draft time: coaching changes, scheme fit, injuries, depth chart competition, and pure variance. The remaining variance is not noise to dismiss; it's a quantified ceiling on what's predictable from pre-draft information.
Calibration — does the model's predicted outcomes match what actually happens, on average? A well-calibrated model has prediction = actual on average within each prediction bucket, even if individual predictions miss. This is independent of R² — you can have a low-R² model that's still well-calibrated.
DMX Prediction vs Actual Outcome — By Position complete-career prospects only

Each point is one prospect with a complete career. The diagonal line is perfect calibration. Points above the line outperformed prediction; below underperformed. For deeper residual analysis — including ranked overperformer/underperformer leaderboards by position — see the Analytics > Residual Analysis tab.

How DMX is built, step by step — this section explains the modeling philosophy without revealing the position-specific weights that make DMX distinct. Anyone could build a composite of ATH/DPOS/AWP; what matters is which one predicts NFL outcomes best, which requires 25 years of fitted refinement.
Component Construction
  1. Athleticism (ATH). Combine and pro-day measurables: 40-yard dash, vertical jump, broad jump, 3-cone, shuttle, bench press, height-adjusted weight. Each metric z-scored within position cohort, then composite-weighted into four sub-components (Speed, Agility, Lower Power, Strength).
  2. Draft Position (DPOS). Pick number transformed to z-score within position cohort. UDFAs assigned pick 300. This single dimension is the strongest predictor at every position — the NFL is collectively pretty good at evaluating prospects.
  3. Age-Weighted Production (AWP). Final college season's per-game production normalized by age and competition tier (FBS Power 5 / G5 / FCS). For IDP, replaced with havoc share to handle scheme variance.
  4. Composite assembly. Per-position weights blend the three components. Weights are calibrated by maximum-likelihood fit against career outcomes on a 25-year training set, refit roughly every 3 years. Specific weight values are not published.
Decile Calibration

Raw DMX is a continuous z-score. The decile assignment (D1–D10) is computed per (position, draft-year) cohort using NTILE(10) OVER (PARTITION BY position, draft_year ORDER BY dmx DESC). This means D1 is always relative to that year's class — a D1 in a weak class is not the same prospect as a D1 in a stacked class. The raw DMX score is more reliable for cross-class comparison.

What's Not in DMX
  • Landing spot. DMX is computed at the moment of the draft itself, so the prospect's team, coaching staff, depth chart situation, and offensive scheme are not encoded.
  • Injury history. Pre-draft injuries are visible to NFL scouts and reflected in DPOS, but not modeled as a separate signal.
  • Character / interview signals. Off-field information is signal-poor and noise-rich; we leave it to the NFL's evaluation expressed through DPOS.
  • Position-versatility flags. A WR/RB hybrid is scored at his primary position. Some signal is lost.
Known biases and data gaps — every model has them. Documented honestly here. None invalidate DMX but every dynasty decision should account for these.
Era Drift

Combine testing standards, college pass-rate context, and draft strategy have all changed materially since 1996. The DMX components are z-scored within draft year to partially absorb this, but the position-weight calibration is fit on long historical windows. A 1998 elite WR profile and a 2024 elite WR profile may look subtly different.

IDP College Data Gaps

Comprehensive defensive college statistics are unavailable in any free, terms-of-service-clean source before 2004. Draft classes 2004 (n=117 prospects) and earlier have AWP set to NULL rather than zero or imputed. These prospects still receive an ATH+DPOS-only DMX, but with lower confidence. The Draft Board surfaces this explicitly.

Small-School and FCS Prospects

AWP component-of-competition adjustments exist but FCS / Division II prospects with limited combine attendance can have noisy ATH. Notable historical examples of dominant FCS prospects (Adrian Peterson at Georgia Southern, the Walter Payton Award winners) are correctly captured, but smaller-school prospects deserve extra qualitative due diligence.

Career-Outcome Definition

"5-year VBD" sums points-above-replacement across years 1–5 post-draft, with a hard window. This treats players with concentrated peaks differently from those with sustained mid-tier seasons. The peak_3yr_vbd_avg regression (also published in the table above) is an alternative formulation that rewards peak performance over total accumulation.

Tier Classification

Career tiers (Star / Starter / Contributor / Bust) are derived from VBD percentiles within position cohort. The boundary thresholds are deterministic but somewhat arbitrary — a player who falls 2% below the Starter threshold is classified Contributor even though their actual career might feel "Starter-adjacent". Use the continuous residual values alongside the tier label.

Trade-Secret Disclosure

Three categories of model internals are not published: (1) DMX position-component weights, (2) IDP DMX position-component weights, and (3) DPX layer weights and dynasty modifier coefficients. These are the result of 25 years of iterative refinement on a curated proprietary dataset and constitute the platform's commercial differentiation. Per Anthropic's trade-secret guidance, exposing only outputs (scores, deciles, rankings, regression coefficients on those outputs) is the appropriate disclosure level. Every value visible in the table above is publicly readable via Supabase. Every value not visible is locked down behind row-level security with anon access revoked.

If you're an independent analyst who wants to validate the model: the regressions above are reproducible. Pull DMX scores for completed-career cohorts, compute career VBD or tier yourself, run OLS, and you should land on coefficients within rounding of what's published here. If you can't, that's a model bug we want to know about.