How the model works

Using Drake London's 2025 actuals (per-17 normalized) in a standard 1 PPR, 0.1 PPY scoring format:

Replacement is the freely-available player at each position — the production a team can plausibly add off waivers. The replacement rank scales with your league's nominal starters plus expected flex usage at that position, with a small buffer for bench-injury insurance. The exact formula varies by position, derived from where each position's production curve actually flattens.

The multipliers are calibrated from empirical cliff analysis on multi-year actuals — each position's production curve has natural drop-off points, and replacement is anchored just past the last meaningful one. TE-specific addition: a hard cliff cap reflects that the talent pool is finite (the empirical TE production curve flattens around rank 30 in standard scoring, ~36 in mild TE-premium, ~42 in heavy TE-premium). Without the cap, large multi-TE leagues would push replacement past where real production exists.

London is 24 — solidly in his prime window. Each future year applies an empirically-derived decay multiplier (small production loss as the player ages) and a cumulative survival probability (the running product of year-over-year retention up to that year). Retention is tier-stratified — London ranks as an elite WR in this league, so he's drawn from the elite-tier retention curve. Tier boundaries are smoothed across a 5-rank band so a one-rank flip at the boundary doesn't produce a mechanical retention cliff; players sitting comfortably inside a tier are unaffected. WR horizon is capped at 9 years where the data still has signal.

Future production is discounted at the league's configured rate (default 8%) per year. WAR is the sum across the position's horizon (WR = 9 years). Both inputs are read directly from your league profile and the position-specific horizon table.

Empirical hit rates re-derived from 2010–2022 NFL draft classes. A "hit" = at least one season above the position-specific production threshold in the player's first 5 NFL years. Bayesian shrinkage (K=10) toward a monotonic positional prior smooths small-sample noise, then strict monotonicity is enforced — a later pick never has a higher hit rate than an earlier one at the same position. Hit probability is the headline diagnostic for a prospect's capital and drives the archetype-signal multipliers below — but it no longer multiplies WAR directly: the outcome curves in step 2 embed hit rate, bust timing, and miss-case production in one empirical object.

The model doesn't assume a prospect either "hits" (and produces a flat median forever) or "busts" (and produces nothing). Instead, for every pick bucket it stores the full realized outcome distribution of the 2010–2022 NFL draft classes, year by NFL year: five representative real seasons (10th / 30th / 50th / 70th / 90th percentile stat lines) — busts included, out-of-league seasons included as zeros, each weighted by the games actually played. Value for each future year is the average of what those five outcomes clear above your league's replacement bar, scored with your league's rates.

Three multiplicative adjustments scale a devy prospect's WAR for the time they're not yet in the NFL. Each captures a distinct concept; together they replace what a single, blunt "discount per year out" would do.

Arrival penalty — ~7% per year discount for non-capital pre-NFL risk: injury, off-field events, landing-spot variance. Draft capital uncertainty is not bundled into this penalty: the model treats your capital input (override or default baseline) as the conviction. If you believe a prospect projects top-5, the model honors that view and tells you what their WAR is if you're right, modulo the small non-capital risks that exist regardless of who's correct about the draft.

Carrying cost — each NFL year of contribution is calendar-discounted at the league's discount rate (typically 8%). This represents the roster-spot opportunity cost: every season you hold a non-producing prospect, you forgo what else that slot could have held. Distinct from arrival penalty, which captures value-disappearance risk. Both apply.

Data confidence taper — the "if you're right" framing assumes a meaningful capital projection exists. That holds for a college senior with three years of film and combine data. It holds less well for a HS recruit who hasn't played a college snap. Top-100 HS recruits get drafted into the NFL at ~35-45% over the next 4-5 years vs ~70% for declared seniors with top-50 projections (Bud Elliott / 247Composite research); mock-to-actual-pick correlation is r≈0.80 in the draft year, r≈0.55 one year out, r≈0.35 two years out. So a 2029 freshman's "projects R1" claim carries roughly half the predictive confidence of the same claim on a 2026 senior. The taper applies a flat multiplier to total devy WAR based on years until NFL entry.

For WRs and TEs with matched college stats, WAR is adjusted by a regression-derived multiplier based on how early the player broke out relative to their draft year (OLS regression across 254 matched WR/TE historical players, 2009–2021 NFL draft classes).

NFL Combine 40-yard time and peak college receiving production are statistically significant predictors of TE fantasy success even after controlling for pick capital. The model applies a credibility multiplier — the empirical "blocker archetype" cell (slow 40 + low college production) historically hits at 5.6% (1 of 18 from 2010–2020) versus 58.3% for the receiving-archetype cell (fast 40 + high college production), a 10× spread the pick-only model misses.

For WRs, the empirical signal is narrower — combine metrics did not discriminate hits from misses (the NFL has already pre-filtered for speed in who gets drafted at all). But for R2–R3 picks specifically, college receiving production matters: WRs with peak ≥1200 single-season college rec yards historically hit at 42% versus 25% for mid-production peers. The model applies a 1.10× lift to that empirical cell — a conservative, lift-only adjustment with no penalty for unknowns.

For RBs, the surprising empirical finding is that combine metrics don't discriminate (40-yard time, weight, vertical, broad — all n.s.) AND college rushing production doesn't discriminate either. What does is college receiving production — the 3-down archetype signal. R2-R3 RBs with peak ≥400 single-season college rec yards historically hit at 71% (n=7) versus 24% for low-rec peers (n=29). Modern fantasy RBs need pass-catching ability — pure rushers (Mark Ingram, Latavius Murray archetype) don't translate even with big college rushing numbers. The model applies a 1.15× lift to the R2-R3 high-rec cell. Small historical sample (n=7) so the lift is conservative versus the empirical ratio (~1.56×).

For QBs, the empirical signal is the inverse of what most people expect. College passing yards do NOT discriminate hits from misses (t=0.82, n.s.), but career college rushing yards strongly do (t=+3.05, p<0.01) — fantasy QB hits averaged 1,077 career college rush yards versus 442 for misses. This is the Konami-code archetype: mobile QBs (Lamar, Allen, Hurts, Daniels) dominate fantasy because rushing stacks on top of passing for 50-100+ extra points per season. Pocket passers need elite efficiency to crack QB1 territory. The model applies a 1.15× lift to picks 33+ with career college rushing ≥1,500 yards. Pick gate skips top-10 (where pure capital dominates — top-10 QBs hit at 60-100% regardless of rushing). Lift-only design — pocket passers are not penalized.

The multipliers above scale hit probability based on signal strength — but pick capital still drives the floor. A pure pick-capital model can still over-project a R2 prospect with confirmed-bad signals (the "Boerkircher case") because R2 capital alone implies ~50% historical hit rate. The cohort cap addresses this by checking the actual empirical hit rate in the player's specific archetype cell.

For every drafted devy prospect, the model looks up their (pick bucket × archetype signal) cell in a pre-computed cohort hit-rate table. If that cell's smoothed historical hit rate is below 10%, AND the model's current effective hit prob exceeds the cell rate, the cap reduces effective hit prob to match the empirical floor. The cap can only decrease — never inflate — and never fires when the cohort has zero observations of that archetype.

Cohort dimensions are position-specific, mirroring each position's empirical battery:

The strongest empirical signal sits at QB: the cohort of R6–R7 pocket-passer QBs (career college rushing < 500 yards) hit at 0 of 16 historically — a smoothed cell rate of ~1%. Those prospects get their effective hit prob capped to match. Other cap-firing cells: R6–R7 high-production WRs (8.9%), R6–R7 mid-production TEs (~2.8%), and R4–R5 pocket QBs (8.5%). All low-magnitude in absolute WAR terms — these are late-round prospects whose pick-capital floor was previously generating "small but non-zero" WAR the empirical record doesn't support.

The cap complements the v3 multipliers rather than replacing them. The multipliers handle continuous signals (a 4.69 forty is meaningfully different from a 4.73 forty). The cap handles discrete cohort empties — cases where the archetype is so empirically rare that even the floor of the multiplicative model overshoots reality. Together they encode the full structure of the empirical record: where signals discriminate, scale; where cohorts confirm low, cap.