A better starting point for player value

Use a veteran’s longer track record, give rookies credit for a plausible developing role, and let new evidence change the estimate.

Research snapshot: 2026-09-13 · 2012–2025 statistics · 86,817 offensive player-games

The result: Power Rankings now uses up to three completed seasons, small fitted age and experience adjustments where supported, and a rookie prior based on draft position and workload. These estimate usefulness during the current season. They do not estimate dynasty trade price or guarantee a breakout.

1. What was missing

The previous historical component used only the preceding season. Within that season, it selected the final 12 appearances and gave them a four-game half-life: a performance eight appearances earlier received one quarter of the latest game’s weight. That is useful for detecting recent role changes, but it can be too reactive when estimating a player’s underlying ability across an offseason.

A rookie had no earlier NFL season to contribute. The model consequently relied on a current weekly projection and whatever few actual games existed. It did not use draft investment, age, or years of experience. This left it without a separate way to recognize development potential when an initial weekly forecast was low.

Simply adding a youth bonus would create a different problem. A young reserve with no credible workload could outrank an established starter. Likewise, replacing a veteran’s recent production with his best career season would systematically overrate past peaks. The changes therefore use weighted histories and groups of comparable players, with explicit limits on their influence.

2. What the outside research supports

Opportunity is essential

Scott Barrett’s weighted-opportunity work distinguishes targets from carries and shows why raw touches lose important information. Its reported correlations are largely descriptive relationships within a season, so they do not prove the accuracy of a new future-season forecast. The useful principle for this model is to retain the type and amount of work a player earns, rather than judging only touchdowns or fantasy points. PFF: Weighted opportunity.

SumerSports’ receiving analysis also distinguishes situation from performance. Route-based efficiency changes with personnel and play context, and a descriptive receiving metric need not be the strongest predictive one. That supports considering role and team context, but it does not justify inventing route counts from offensive snaps. This implementation keeps snaps, targets, and routes conceptually separate. SumerSports: Revisiting yards per route run.

Age patterns depend on position and sample selection

Chase Stuart’s running-back study compares adjacent seasons rather than only complete careers. It explicitly discusses the survivorship problem: requiring substantial carries in both seasons omits players whose workload disappeared. That is a useful warning against applying a curve estimated only from the stars who lasted. Our comparison retains returning players even when their workload collapses and includes a separate check assigning zero to nonreturners. Football Perspective: Running-back aging patterns.

Neil Paine’s quarterback analysis examines changes in era-adjusted passing efficiency. This reinforces treating passing performance separately from simple age-based labels, but its efficiency curve is not a fantasy-points coefficient. We test age and rushing contribution directly on our own historical sample instead of copying its curve. Football Perspective: Quarterback aging.

Apex’s quarterback study finds different age distributions among high-scoring pocket passers and runners. Its sample is selected for strong fantasy seasons; the distribution of those successful seasons cannot establish an individual player’s probability of decline. It motivated testing an interaction between age and rushing contribution, rather than imposing a single quarterback age cliff. Apex: Quarterback peak-age study.

Apex’s tight-end study shows how a handful of exceptionally productive players affect the older end of the distribution. Its receiver study also examines changes across eras. Those are useful reasons to distinguish positions and limit demographic adjustments. They do not support automatically discounting every older player or boosting every young one. Apex: Tight ends · Apex: Wide receivers.

Draft pedigree provides a prior, not a verdict

Apex’s rookie-receiver study describes the draft positions, entry ages, and college profiles of successful rookies. Early draft selections are common in that group. However, the share of successful players who were early picks is not the same as the share of early picks who succeeded. To estimate a useful prior, our study includes the broader rookie population, not just named breakout receivers. Apex: Rookie receiver breakouts.

The practical inference is that pedigree can improve an initial estimate when NFL evidence is scarce, but its influence should decline as actual workload and production accumulate. Strong NFL evidence must be able to override a late draft selection. No player gets a permanent draft-round bonus.

3. Data collected and prepared

The study downloaded 14 regular-season weekly-stat datasets, player identifiers and biographical information, draft records, and available snap-count files from nflverse. Its player-stat documentation describes box-score-aligned outputs; its player metadata documents birthdates, rookie seasons, draft information, and cross-provider IDs. nflverse player stats · Player metadata · Draft records.

Postseason games are excluded. Snap records are joined through player IDs, and duplicate player/week records are rejected. Offensive appearances with no target or carry remain in the sample when snap data establishes participation. An active-list entry alone is not a played game. The downloaded 2012 snap file was empty, so that year relies on recorded statistical opportunities; usable snap coverage begins in 2013. nflverse snap-count documentation.

Age is calculated at September 1 of the season being predicted. Experience is calculated from rookie season, rather than using a player’s present age or career length in historical forecasts. A model predicting 2024 therefore cannot use the player’s 2025 performance. Current production forecasts use only previously completed seasons and settled current-season results through the displayed week.

The saved runtime snapshot contains 1,233 offensive player profiles, their available 2023–2025 season averages, draft and biographical fields, and fitted parameters. Large research inputs and reproducibility records stay separate from the website’s initial download. The model joins ESPN IDs to existing player identities and verifies position; matching a player’s name alone is insufficient to attach a career.

4. The historical comparison

For established players, model development uses target seasons 2015–2021. The 2022–2023 seasons select annual history decay and the strength of a small regularized demographic correction. The table below evaluates 2024–2025 using coefficients fitted only through 2023. After evaluation, the chosen configuration is refitted through 2025 for use in 2026.

Eligible players needed at least four prior-season appearances and a modest prior workload: 10 pass attempts per appearance for QBs, three carries plus targets for RBs, or one target for WRs and TEs. No minimum future workload is required. The main target is next-season fantasy points per offensive appearance among returners. A separate nonreturner sensitivity check preserves players with no subsequent appearance and assigns zero.

Every position preferred annual decay of 0.25 among the tested values. With equal, sufficiently large season samples, that gives roughly 76% of the career component to the latest year, 19% to the year before, and 5% to the third year. Each season’s evidence is capped at 12 appearances for weighting. A short season contributes less evidence; it is not treated as zero production for missed games.

Age and experience corrections improved the selected QB, RB, and WR configuration. Validation selected zero demographic adjustment for TEs. Corrections are limited to ±20% of the career component; they are never multiplied over the entire live forecast. Because the fitted intercept also corrects average overprediction, the displayed adjustment should not be interpreted as a causal biological effect of age.

The new career component also improves the aggregate error comparison when nonreturners receive zero. That helps check survivorship sensitivity, but it does not turn this into a model of injury recovery or availability. Some subgroups still worsen: for example, the down-year TE subgroup’s absolute error does not improve. A favorable aggregate comparison is not a promise about every position subgroup or player.

Scope of the evidence: these are tests of historical inputs. The “recent 12-game” comparator reproduces the old history window and decay using raw fantasy scoring; it is not the full v1 model with its published projections and touchdown treatment. Archived pregame forecasts are unavailable across these seasons. These results must not be presented as a measured improvement in the accuracy of the complete live site.

5. Rookie development and speculative value

A preliminary draft-only prior reduced some large misses but increased average absolute error, especially when it assigned positive value to rookies with no established work. The revised approach conditions on actual role. Historical rookie checkpoints after Weeks 1, 4, and 8 predict the remainder of that rookie season, including zero for players who make no later appearance.

Draft groups are picks 1–32, picks 33–100, later selections, and unavailable draft position. Missing draft information is not relabeled as a confirmed undrafted player. Workload groups use pass attempts for QBs, carries plus targets for RBs, and targets for receivers. The low/regular/high boundaries are 10/25 for QBs, 5/15 for RBs, 2/6 for WRs, and 2/5 for TEs.

Each cell is blended toward the broader position-and-role average using ten equivalent players. The cell’s confidence uses its number of distinct players, rather than treating repeated checkpoints for the same rookie as independent people. The expected remaining-season stat line is then scored using the league’s settings, including reception and TE premiums.

The revised rookie comparison improves both error measures for QB, WR, and TE. RB results are mixed: it reduces large errors, as measured by RMSE, but increases typical absolute error. The live model therefore gives the RB rookie component half the maximum weight used for the other positions. That weight reduction is a conservative policy choice, not a separately validated optimum.

Rookie role conditioning was refined after looking at aggregate retrospective diagnostics. These results are development-informed historical comparisons, not an untouched prospective trial. Checkpoints from the same rookie are correlated, and the QB/TE samples are small. The report deliberately does not assign a statistical confidence interval or a personalized breakout probability.

In the site, a positive forecast or observed offensive workload is required. A healthy zero projection, no credible workload information, or a player explicitly without an NFL team receives no extra rookie component. If no current actual role is available, projected attempts/targets provide the role estimate; that preseason use is an inference from the historical relationship, not a separately tested preseason-forecast result.

The new component is modest, fades as games accumulate, and cannot use an upward estimate above 1.5 times a positive weekly forecast plus one point. These limits protect against double-counting expectations already present in published projections. Rookies can move either up or down. The model values their plausible production rather than adding their best imaginable outcome to their expected score.

6. What this means for Lamar Jackson

The concern about overreacting to his last season is supported by the input structure. His official career page shows substantially stronger passing and rushing production in 2023 and 2024 than in 2025. Under the study’s four-point passing-TD PPR scoring, including lost fumbles and two-point conversions, the corresponding historical averages are approximately 20.70, 25.32, and 16.53 points per offensive appearance. Baltimore Ravens: Lamar Jackson’s career statistics.

The old recency window reduces the raw 2025 profile further, to about 14.09 before its other adjustments. The new career component retains all three seasons with declining annual weight. That restores some evidence of his established level without replacing the most recent season with his career peak. His age, experience, and rushing contribution follow the same fitted rules as every other QB. There is no Lamar-specific floor, multiplier, or forced positional rank.

A genuine sustained loss of role should still lower a player’s value. Current projections and recent usage remain substantial inputs, and a healthy forecast far below a player’s career baseline reduces the weight of that historical baseline. New information can therefore outweigh old success.

7. Signals considered but not added as separate bonuses

College production, breakout age, early declaration, and athletic testing: potentially useful before a player has NFL evidence, but no consistently verified, player-matched college dataset was available in this implementation. Draft position is an imperfect summary of prospect evaluation. Adding manually collected statistics for a few famous rookies would make coverage uneven; these require a broader dataset and a new time-separated test.

Routes, target earning, air yards, and valuable field position: good candidates for a richer opportunity model. The available current data does not establish consistent routes or red-zone work for every player. Offensive snap share cannot substitute for route participation. Public weekly statistics alone do not solve that missing-information problem.

Team strength, coaching changes, and depth-chart movement: these affect future opportunity, but the existing weekly forecasts and observed workload already reflect some of that context. A separate team-win bonus could count the same information twice. A future improvement should estimate opportunity changes explicitly and test them against a projection-only benchmark.

Injury history and age-related availability: per-appearance production and probability of being available are different quantities. The current status discounts remain visible assumptions. This update does not infer diagnoses, fabricate recovery dates, or convert historical injuries into confident individual forecasts.

Dynasty value and upside: a long career horizon would value future seasons and trade liquidity differently. A backup’s breakout upside can matter to bench construction, but adding a ceiling directly to every player’s mean would systematically overvalue uncertainty. A later dynasty or scenario view should show separate horizons and probabilities with appropriate validation.

8. Reuse, performance, and next validation

The maintained source contains the research script, source URLs and checksums, calibration results, and the compact runtime dataset. This lets later changes reuse the historical work rather than starting from scattered articles. The website loads the versioned research asset once when Power Rankings needs it and shares it across leagues. Its raw size is about 399 KB, or roughly 82 KB under gzip; it is not embedded in the main HTML or fetched during login.

Supabase continues to supply current projections, settled actual results, identities, and league state. A complete saved historical season fills a gap when a live history response has fewer appearances; the same season is never counted twice. If the research asset is unavailable, the page retains its available projections and season history and displays a notice. Kicker and defense models, team-lineup assignment, scoring conventions, and roster depth caps retain their existing behavior.

The most useful next experiment is a timestamped archive of future live forecasts and inputs. That would let us compare the complete v2 model against the complete previous model, broken out by position, experience, availability, and week. Until that exists, the research supports improved historical reasoning and more explicit uncertainty—not a claim of proven live forecasting superiority.