Every combine number, college production line, draft slot, and career outcome from the last 25 years, linked player by player. This is the workbench where I do the digging.
3,517
Scored Prospects
1,967
Outcomes Linked
27
Draft Classes
12,730
VBD Records
10
Analysis Modules
Player Compare — stack 2–3 guys side by side on their DMX (pre-draft prospect score) or DPX (in-career dynasty value) component z-scores. Offense and IDP both work. I built this one to screenshot, so hit Download image and you get a clean, credited graphic to drop in your league chat.
Pre-draft prospect score — offense & IDP
Select players to compare
Add at least two players above.
The image carries a dynastyfootballfan.com credit, so it's yours to post anywhere.
Hit Rate Matrix — of all the prospects who scored in each DMX decile, how many actually put up a Star or Starter season? That's the whole question. Star/Starter (the default) means a guy cleared a peak-season bar inside the eval window: QB ≥200 PPR pts, RB/WR ≥120, TE ≥75. Stars have to clear a higher bar (QB 300, RB/WR 200, TE 150). The window is 5 yrs for QB/RB/WR and 7 yrs for TE, since tight ends take longer to come around. Every cell shows the hit rate, and if you hover you get the Wilson 95% confidence interval — the honest margin given how many guys sit in that decile. Cohort is 2001–2020 (TE through 2018). Flip the metric to Star Only, Top-20% PPG (the old percentile-based definition), Opportunity (share who ever got NFL snaps), or Conditional (the Star/Starter rate among the guys who actually played). One caveat I'll own: those Star/Starter bars are absolute PPR points, not era-adjusted, so they're marked for future review. Want the model checked out-of-sample? That's the Model Performance tab.
Hit Rate by DecileRB/WR/QB: 5-yr window · TE: 7-yr window · 2001–2020 classes
Metric:% with a Star or Starter peak season (QB ≥200, RB/WR ≥120, TE ≥75 PPR pts)
Loading…
Offense Curveshit rate by decile · TE shows 7-yr window
IDP Curveshit rate by decile
Model Performance — here's the honest test: how well does DMX call NFL careers when I validate it out-of-sample? Every draft class gets predicted from training data strictly before it (time-series cross-validation, expanding window), so the model never peeks at the answer. Test cohort is 2011–2020 prospects (2011–2018 for TE). For each (position, decile), the predicted Star/Starter probability is just the historical hit rate in the earlier classes, and the predicted tier is the historical mean tier rounded off. For the empirical hit-rate lift by decile, go to the Hit Rate Matrix tab. This tab asks one question and only one — is the model honest about itself?
Position
Out-of-Sample PerformanceTime-series cross-validation, all positions
Calibration Plotpredicted vs observed Star/Starter rate
Each dot is a DMX decile from the held-out test set. A dot above the 45° line means the observed hit rate beat what the training era predicted; below the line, it came up short. The vertical bars are 95% confidence intervals on the observed rate. When the dots drift above the line as a group, that's an era effect — modern hit rates run higher than the training-era rates did.
Rows are what the model predicted; columns are what actually happened. Each cell is the share of that predicted-tier row that landed in each actual tier, so the diagonal is where the model got it right. Cohen’s κ is the chance-corrected agreement rate (0 = random, 1 = perfect). You'll see Bust–Contributor and Starter–Star mixups a lot. The model nails the extremes; it's the tiers right next to each other that blur.
Methodology & Limitations
Time-series cross-validation (TSCV): for every prospect in the test set (draft years 2011–2020), I predict the outcome using only training data from earlier draft years. The expanding window mimics how I'd actually have used the model in real time, and it gives the most honest out-of-sample read you can get.
Why I predict by decile: the (position, decile) bucket is the canonical DMX grouping I use everywhere on the site. Predicting a bucket's historical hit rate sidesteps the parametric assumptions (linearity, normality) that just don’t hold for censored 5-yr VBD outcomes. It’s the same move a statistician would make for empirical-Bayes calibration.
Brier skill score: stacks the model’s Brier score (mean squared error of predicted probability vs observed outcome) against a dumb baseline that just predicts the overall hit rate every time. Positive means the model earns its keep beyond the base rate, and the bigger the number, the more it earns.
Cohen’s κ interpretation (Landis & Koch 1977): <0.20 slight, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 substantial, >0.80 near-perfect. Calling tiers this far out — from one composite score, for a five-year outcome — lands in the fair-to-moderate range. For a problem this hard, that’s about what you should expect.
Limitations:
Era effect: hit rates roughly doubled from 2001–2010 to 2011–2020 as offense took off league-wide. Anything trained on the older data will under-predict the modern hit rates, and you can see it as upward drift on the calibration plot.
Sample size: QB deciles run about ~12 prospects apiece, so the confidence intervals get wide. Whatever I say about QB rests on smaller cohorts than WR or RB.
Star/Starter thresholds are absolute (QB 200, RB/WR 120, TE 75 PPR pts). Era-adjusted bars (position rank within the season) would be more rigorous, and I've got that marked for future review.
Ceiling on predictability: 60–80% of what happens to a career comes from stuff that lands after draft day (coaching, scheme, injury, QB play, team success), none of which a pre-draft model can see. Brier skill scores in the 0.08–0.27 range fit right under that ceiling.
Draft Capital Efficiency — VBD ROI broken out by draft round. First-round RBs (picks 1–10) put up 128.0 average 5-year VBD at a 78.8% hit rate, and by both numbers that's the best dynasty bet on the board. Round 2-3 QBs? A 2.1% star rate on 48 prospects. About as bad as it gets.
78.8%
RB Rd1 Hit Rate
picks 1-10 · n=118
128.0
RB Rd1 Avg VBD
5-yr cumulative PPR
66.3%
WR Rd1 Hit Rate
picks 1-10 · n=205
2.1%
QB Rd2-3 Star Rate
1 star from 48 prospects
Metric
Draft Capital Efficiency by Position & Round Tier
Full BreakdownVBD · Hit% · Star%
Positional Scarcity Waves — the share of guys still above replacement-level fantasy value, year by year after the draft. Your sell windows live in the slope of each curve. RBs cliff. WRs plateau. TEs are a slow burn. QBs are all-or-nothing.
PositionDeciles
% Above ReplacementRB
Avg VBD by Year Post-Draft
Position Insights
Select position and click Update.
Player Profiles — a descriptive archetype taxonomy built from k-means clustering on DMX-orthogonal residuals. These profiles tell you how a prospect earned his score (playing style, college role); they don't try to predict his career. For that, lean on the DMX and DPX deciles. RBs sort into 4 profiles, Edge defenders into 3. Beta — v1.0
Position group
RBEdge
Find a player
—
Profiles in group
—
Prospects classified
—
Largest profile
—
Features per player
Loading profiles…
Methodologywhy these are descriptive, not predictive
Profiles are k-means cluster assignments run on DMX-orthogonal residuals — I strip the DMX score axis out mathematically before clustering, so what's left describes shape and makeup, not overall quality. Two guys with the same DMX can land in different profiles, and two guys in the same profile can carry very different DMX scores.
I set the cluster counts (k) empirically per position group with silhouette analysis — 4 for RB (n=158 cluster-ready) and 3 for Edge (n=125). When I checked those clusters against career outcomes, they read as interpretable but they do not separate on hit-rate at p<0.05. That's exactly why I treat them as taxonomy and not prediction.
Each player gets assigned by centroid distance in residual space, but only if he clears a minimum-features bar (7 of 9 for RB, 5 of 7 for Edge). If a guy is missing too many, I leave him unassigned on purpose instead of slapping the wrong label on him.
Cross-Era Analysis — how has DMX shifted from one draft class to the next? Watch the D1 threshold, the class makeup, and how the components drift from 2000 to 2026. You can also put any two classes head-to-head on their ATH/DPOS/AWP profiles.
PositionComponent
WR Avg DMX by Draft Year
Class Comparison Toolside-by-side ATH · DPOS · AWP radar
Class Avs Class B
Combine-to-Career Regression — once you control for draft capital, which combine tests actually predict NFL success? Here's what the data says. DPOS dominates every individual test at every position. The 40 time is the most overrated number of the bunch. Broad jump is the best single athleticism read for QBs and TEs. And AWP (college production) beats every combine test there is except DPOS.
~1.5%
ATH Alone R²
all positions — athleticism barely predicts
-0.25
TE 40-Time Corr
best single combine test correlation
0.202
QB Broad Jump Corr
best combine test for quarterbacks
13.4%
RB DPOS R²
draft capital vs VBD — strongest signal
View
Combine Test Correlation with 5-Yr VBDby position
AWP Predictive Premiumstandalone R² vs ATH vs DPOS vs DMX
Key Insight
DPOS (the draft-capital z-score) is the most predictive variable at every position, explaining 16–22% of 5-year VBD variance all by itself. AWP (college production) chips in another 2.6–9.1% of standalone signal, which makes it the best predictor you can get that has nothing to do with athleticism. ATH on its own? Just 1.7–5.4%. Once you pull raw combine athleticism away from draft position and college context, it barely moves the needle. The 40-yard dash is the worst offender, explaining near-zero variance for WRs and RBs. Broad jump and vertical hold up better as single tests at most spots. And the DMX composite beats any one test by design, which was the whole idea — this data just confirms I built it the right way.
Distribution Analysis — how DMX scores spread out within and across the deciles, so you can see how cleanly the model pulls the quality tiers apart. There are individual ATH/DPOS/AWP radar profiles in here too for scouting context — pick any draft year and you get the full class laid out.
PositionViewYear (Radar)
WR DMX Distribution by Decile
Draft Board Heat MapATH · DPOS · AWP strength by draft class
Interactive Scatter — plot any DMX component pair against career outcomes. Hover to pull up the player and his stats; the color is his career tier. It's the quickest way to spot where the signal is strong and where it breaks down completely.
PositionX AxisY Axis
DMX vs 5-Yr VBD — WRhover for player · color = career tier
StarStarterContributorBust
AI Analytics Assistant — ask questions in plain English, grounded in the real DFF dataset. It gets the full context automatically — the DMX/DPX methodology, the hit rates, the R² values, the combine regressions, the draft-capital efficiency, and the residual-analysis findings. Anything you ask gets saved to your library so you can come back to it.
Natural Language AnalyticsPowered by Claude
Cmd+Enter to submit
Saved Query Library
No saved queries yet.
Context: DMX formula, all position hit rates, R² values, combine correlations, draft capital data, residual analysis, 25-yr historical scope.